SAP PI

f

SAP NetWeaver Process Integration (PI) Performance Check Guide: Analyzing Performance Problems and Possible Solution Strategies SAP NetWeaver PI 7.1x SAP NetWeaver PI 7.3x SAP NetWeaver PI 7.4

Document Version 3.0 March 2014

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

DOCUMENT HISTORY Before you start planning, make sure you have the latest version of this document. You can find the link to the latest version via SAP Note 894509 – PI Performance Check. The following table provides an overview on the most important changes to this document. Version

Date

Description

3.0

March 2014

Reason for new version: Include latest improvements in PI 7.31 and PI 7.4 and enhanced tuning options for Java only runtime. Many sections were updated. Major changes are:

 Additional symptom (blacklisting) for qRFC queues in READY status

 Long processing time for Lean Message Search  IDoc posting configuration on receiver side  Tuning of SOAP sender adapter  More information about IDOC_AAE adapter tuning  Packaging for Java IDoc and Java Proxy adapter  New Adapter Framework Scheduler for polling adapters  new performance monitor for Adapter Engine  Enhancements of the maxReceiver parameter for individual interfaces

 Include latest information for staging and logging (restructured to be part of Java only chapter)

 FCA Thread tuning for incoming HTTP requests  Enhancements in the Proxy framework on sender/receiver ERPs  Large message queues on PI Adapter Engine  Generic J2EE database monitoring in NWA  Appendix section: XPI_Inspector 2.0

October 2011

Reason for new version: Adaption to PI 7.3

 Added document history  Added options and description about AEX  Added new adapters / updated existing adapter parallelization  Description of additional monitors (Java performance monitor and new ccBPM monitor)

 General review of entire document with corrections  More details and performance measurements for Java Only 2

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

(ICO) scenarios

Additional new chapters:

 Prevent blocking of EO queues  Avoid uneven backlogs with queue balancing  Reduce the number of EOIO queues  Adapter Framework Scheduler  Avoid blocking of Java only scenarios  J2EE HTTP load balancing  Persistence of Audit Log information in PI 7.10 and higher  Logging / Staging on the AAE (PI 7.3 and higher) 1.1

December 2009

Reason for new version: Adaption to PI 7.1

 General review of entire document with corrections Additional new chapters:

 Tuning the IDoc adapter  Message Prioritization on the ABAP stack  Prioritization in the Messaging System  Avoid blocking caused by single slow/hanging receiver interface  Performance of Module Processing  Advanced Adapter Engine / Integrated Configuration  ABAP Proxy system tuning  Wily Transaction Trace 1.0

October 2007

Initial document version released for XI 3.0 and PI 7.0

3

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

TABLE OF CONTENTS 1

INTRODUCTION ............................................................................................................................... 6

2

WORKING WITH THIS DOCUMENT .............................................................................................. 10

3 3.1 3.2 3.2.1 3.3

DETERMINING THE BOTTLENECK .............................................................................................. 12 Integration Engine Processing Time ........................................................................................... 12 Adapter Engine Processing Time ................................................................................................ 13 Adapter Engine Performance monitor in PI 7.31 and higher ........................................................... 15 Processing Time in the Business Process Engine .................................................................... 16

4 ANALYZING THE INTEGRATION ENGINE ................................................................................... 20 4.1 Work Process Overview (SM50/SM66) ........................................................................................ 21 4.2 qRFC Resources (SARFC) ............................................................................................................ 22 4.3 Parallelization of PI qRFC Queues ............................................................................................... 23 4.4 Analyzing the runtime of PI pipeline steps ................................................................................. 29 4.4.1 Long Processing Times for “PLSRV_RECEIVER_ DETERMINATION” / PLSRV_INTERFACE_DETERMINATION....................................................................................................... 30 4.4.2 Long Processing Times for “PLSRV_MAPPING_REQUEST” ........................................................ 30 4.4.3 Long Processing Times for “PLSRV_CALL_ADAPTER” ................................................................ 33 4.4.4 Long Processing Times for “DB_ENTRY_QUEUEING” .................................................................. 34 4.4.5 Long Processing Times for “DB_SPLITTER_QUEUEING” ............................................................. 34 4.4.6 Long Processing Times for “LMS_EXTRACTION” .......................................................................... 36 4.4.7 Other step performed in the ABAP pipeline..................................................................................... 36 4.5 PI Message Packaging for Integration Engine............................................................................ 37 4.6 Prevent blocking of EO queues ................................................................................................... 38 4.7 Avoid uneven backlogs with queue balancing ........................................................................... 38 4.8 Reduce the number of parallel EOIO queues ............................................................................. 40 4.9 Tuning the ABAP IDoc Adapter .................................................................................................... 41 4.9.1 ABAP basis tuning ........................................................................................................................... 41 4.9.2 Packaging on sender and receiver side .......................................................................................... 42 4.9.3 Configuration of IDoc posting on receiver side ................................................................................ 42 4.10 Message Prioritization on the ABAP Stack ................................................................................. 44 5 5.1 5.2 5.3 5.4 5.5 5.6

ANALYZING THE BUSINESS PROCESS ENGINE (BPE) ............................................................ 45 Work Process Overview ................................................................................................................ 45 Duration of Integration Process Steps ........................................................................................ 45 Advanced Analysis: Load Created by the Business Process Engine (ST03N) ....................... 47 Database Reorganization .............................................................................................................. 47 Queuing and ccBPM (or increasing Parallelism for ccBPM Processes) ................................. 48 Message Packaging in BPE .......................................................................................................... 48

6 6.1 6.1.1 6.1.2 6.1.3 6.1.4 6.1.5 6.1.6 6.2 6.2.1 6.2.2 6.2.3 6.2.4 6.2.5 6.3 6.4

ANALYZING THE (ADVANCED) ADAPTER ENGINE .................................................................. 50 Adapter Performance Problem ..................................................................................................... 51 Adapter Parallelism.......................................................................................................................... 51 Sender Adapter ................................................................................................................................ 54 Receiver Adapter ............................................................................................................................. 55 IDoc_AAE adapter tuning ................................................................................................................ 56 Packaging for Proxy (SOAP in XI 3.0 protocol) and Java IDoc adapter ......................................... 56 Adapter Framework Scheduler ........................................................................................................ 58 Messaging System Bottleneck ..................................................................................................... 60 Messaging System in between AFW Sender Adapter and Integration Server (Outbound) ............ 63 Messaging System in Between Integration Server and AFW Receiver Adapter (Inbound) ............ 64 Interface Prioritization in the Messaging System............................................................................. 64 Avoid Blocking Caused by Single Slow/Hanging Receiver Interface .............................................. 65 Overhead based on interface pattern being used ........................................................................... 67 Performance of Module Processing ............................................................................................ 68 Java only scenarios / Integrated Configuration objects ............................................................ 69 4

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

6.4.1 6.4.2 6.4.3 6.4.4 6.5 6.6 6.6.1 6.6.2 6.6.3 6.6.4

General performance gain when using Java only scenarios ........................................................... 69 Message Flow of Java only scenarios ............................................................................................. 71 Avoid blocking of Java only scenarios ............................................................................................. 73 Logging / Staging on the AAE (PI 7.3 and higher) .......................................................................... 73 J2EE HTTP load balancing ........................................................................................................... 75 J2EE Engine Bottleneck................................................................................................................ 76 Java Memory ................................................................................................................................... 76 Java System and Application Threads ............................................................................................ 80 FCA Server Threads ........................................................................................................................ 83 Switch Off VMC ............................................................................................................................... 84

7 7.1

ABAP PROXY SYSTEM TUNING .................................................................................................. 85 New enhancements in Proxy queuing ......................................................................................... 86

8 8.1 8.2

MESSAGE SIZE AS SOURCE OF PERFORMANCE PROBLEMS .............................................. 87 Large message queues on PI ABAP ............................................................................................ 88 Large message queues on PI Adapter Engine ........................................................................... 88

9 9.1 9.2 9.3 9.3.1 9.3.2 9.3.3 9.3.4 9.3.5 9.4

GENERAL HARDWARE BOTTLENECK ....................................................................................... 90 Monitoring CPU Capacity.............................................................................................................. 90 Monitoring Memory and Paging Activity ..................................................................................... 91 Monitoring the Database ............................................................................................................... 91 Generic J2EE database monitoring in NWA.................................................................................... 92 Monitoring Database (Oracle) ......................................................................................................... 93 Monitoring Database (MS SQL) ...................................................................................................... 94 Monitoring Database (DB2) ............................................................................................................. 95 Monitoring Database (MaxDB / SAP DB) ........................................................................................ 97 Monitoring Database Tables ......................................................................................................... 99

10 10.1 10.2 10.3 10.3.1

TRACES, LOGS, AND MONITORING DECREASING PERFORMANCE ................................... 101 Integration Engine ....................................................................................................................... 101 Business Process Engine ........................................................................................................... 102 Adapter Framework ..................................................................................................................... 102 Persistence of Audit Log information in PI 7.10 and higher .......................................................... 103

11

ERRORS AS SOURCE OF PERFORMANCE PROBLEMS ........................................................ 104

APPENDIX A ................................................................................................................................................. 105 A.1 Wily Introscope Transaction Trace ............................................................................................ 105 A.2 XPI inspector for troubleshooting and performance analysis ................................................ 106

5

SAP NETWEAVER 7.1 AND HIGHER

1

PERFORMANCE CHECK GUIDE

INTRODUCTION

SAP NetWeaver Process Integration (PI) consists of the following functional components: Integration Builder (including Enterprise Service Repository (ESR), Service Registry (SR) and Integration Directory), Integration Server (including Integration Engine, Business Process Engine and Adapter Engine), Runtime Workbench (RWB), and System Landscape Directory (SLD). The SLD, in contrast to the other components, may not necessarily be part of the PI system in your system landscape because it is used by several SAP NetWeaver products (however it will be accessed by the PI system regularly). Additional components in your PI landscape might be the Partner Connectivity Kit (PCK), a J2SE Adapter Engine, one or several non-central Advanced Adapter Engines or an Advanced Adapter Engine Extended. An overview is given in the graphic below. The communication and accessibility of these components can be checked using the PI Readiness Check (SAP Note 817920).

Processing at runtime is carried out by the Integration Server (IS) with the aid of the SLD. The IS in this graphic stands for the Web Application Server 7.0 or higher. In the classic environment PI is installed as a double stack: an ABAP and a Java stack (J2EE Engine). The ABAP stack is responsible for pipeline processing in the Integration Engine (Receiver Determination, Interface Determination, and so on) and the processing of integration processes in the Business Process Engine. Every message has to pass through the ABAP stack. The Java stack executes the mapping of messages (with the exception of ABAP mappings) and hosts the Adapter Framework (AFW). The Adapter Framework contains all XI adapters except the plain HTTP, WSRM, and IDoc adapters.

6

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

With 7.1 the so called Advanced Adapter Engine (AAE) was introduced. The Adapter Engine was enhanced to provide additional routing and mapping call functionality so that it is possible to process messages locally. This means that a message that is handled by a sender and receiver adapter based on J2EE does not need to be forwarded to the ABAP Integration Engine. This saves a lot of internal communication, reduces the response time, and significantly increases the overall throughput. The deployment options and the message flow for 7.1 based systems and higher are shown below: Currently not all the functionalities available in the PI ABAP stack are available on the AAE but each new PI release closes the gap further.

In SAP PI 7.3 and higher the Adapter Engine Extended (AEX) was introduced. In addition to the AAE the Adapter Engine Extended also allows local configuration of the PI objects. The AEX can therefore be seen as a complete PI installation running on Java only. From the runtime perspective no major differences can be seen compared to the AAE and therefore no differentiation is made in this guide.

7

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Looking at the different involved components we get a first impression of where a performance problem might be located: In the Integration Engine itself, in the Business Process Engine, or in one of the Advanced Adapter Engines (central, non-central, plain, PCK). Of course, a performance problem could also occur anywhere in between, for example, in the network or around a firewall. Note that there is a separation between a performance issue “in PI” and a performance issue in any attached systems: If viewed “through the eyes” of a message, (that is, from the point of view of the message flow), then the PI system technically starts as soon as the message enters an adapter or (if HTTP communication is used) as soon as the message enters the pipeline of the Integration Engine. Any delay prior to this must be handled by the sending system. The PI system technically ends as soon as the message reaches the target system, for example, a given receiver adapter (or in case of HTTP communication, the pipeline of the Integration Server) has received the success message of the receiver. Any delay after this point in time must be analyzed in the receiving system. There are generally two types of performance problems. The first is a more general statement that the PI system is slow and has a low performance level, or does not reach the expected throughput. The second is typically connected to a specific interface failing to meet the business expectation with regard to the processing time. The layout of this check is based on the latter: First you should try to determine the component that is responsible for the long processing time or the component that needs the highest absolute time for processing. Once this has been clarified, there is a set of transactions that will help you to analyze the origin of the performance problem. If the recommendations given in this guide are not sufficient, SAP can help you to optimize the service via SAP consulting or SAP MaxAttention services. SAP might also handle smaller problems, restricted to a specific interface, if you describe your problem in a SAP customer incident. Support offerings like “SAP Going Live Analysis (GA)” and “SAP Going Live Verification (GV)” can be used to ensure that the main system parameters are configured according to SAP best practice. You can order any type of service using SAP Service Marketplace or your local SAP contacts. If you have already worked with the PI Admin Check (SAP Note 884865) you will recognize some of the transactions. This is because SAP considers the regular checking of the performance to be an important administrational task. However, this check tries to show the methodology to approach performance problems to its reader. Also, it offers the most common reasons for performance problems and links to possible followup actions but does not refer to any regular administrational transactions.

8

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Wily Introscope is a prerequisite for analyzing performance problems at the Java side. It allows you to monitor the resource usage of multiple J2EE server nodes and provides information about all the important components on the Java stack, like mapping runtime, messaging system, or module processor. Furthermore, the data collected by Wily is stored for several days so it is still possible to analyze a performance problem at a later date. Wily Introscope is provided free-of-charge for SAP customers. It is delivered as part of Solution Manager Diagnostics but can also be installed separately. For more information, see http://service.sap.com/diagnostics or SAP Note 797147.

9

SAP NETWEAVER 7.1 AND HIGHER

2

PERFORMANCE CHECK GUIDE

WORKING WITH THIS DOCUMENT

If you are experiencing performance problems on your Process Integration system, start with Chapter 3: Determining the Bottleneck. It helps you to determine the processing time for the different runtimes within PI Integration Engine, Business Process Engine and Adapter Engine or connected Proxy System. Once you have identified the area in which the bottleneck is most probably located, continue with the relevant chapter. This is not always easy to do because a long processing time is not always an indication for a bottleneck: It can make sense to do this if a complicated step is involved (Business Process Engine), if an extensive mapping is to be executed (IS), or if the payload is quite large (all runtimes). For this reason you need to compare the value that you retrieve from Chapter 3 with values that you have received previously, for example. The history data provided for the Java components by Wily Introscope is a big help. If this is not possible, you will have to work with a hypothesis. Once the area of concern has been identified (or your first assumption leads you there), Chapters 4, 5, 6, and 7 will help you to analyze the Integration Engine, the Business Process Engine, the PI Adapter Framework and the ABAP Proxy runtime. After (or preferably during) the analysis of the different process components, it is important to keep in mind that the bottlenecks you have observed could also be caused by other interfaces processing at the same time. It will lead you to slightly different conclusions and tuning measures. You therefore have to distinguish the following cases based on where the problem occurs: A. with regard to the interface itself, that is, it occurs even if a single message of this interface is processed B. or with regard to the volume of this interface, that is, many messages of this interface are processed at the same time C. or with regard to the overall message volume processed on PI. This can be analyzed by repeating the procedure of Chapter 3 for A) a single message that is processed while no other messages of this interface are processed and while no other interfaces are running. Then compare this value with B) the processing time for a typical amount of messages of this interface, not simply one message as before. If the values of measurement A) and B) are similar, repeat the procedure with C) a typical volume of all interfaces, that is, during a representative timeframe on your productive PI system or with the help of a tailored volume test. These three measurements – A) processing time of a single message, B) processing time of a typical amount of messages of a single interface and C) processing time of a typical amount of messages of all interfaces – should enable you to distinguish between the three possible situations: o

A specific interface has long processing steps that need to be identified and improved. The tuning options are usually limited and a re-design of the interface might be required.

o

The mass processing of an interface leads to high processing times. This situation typically calls for tuning measures.

o

The long processing time is a result of the overall load on the system. This situation can be solved by tuning measures and by taking advantage of PI features, for example, to establish separation of interfaces. If the bottleneck is hardware-related it could also require a re-sizing of the hardware that is used.

10

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Chapters 8, 9, 10 and 11 deal with more general reasons for bad performance, such as heavy tracing/logging, error situations, and general hardware problems. They should be taken into account if the reason for slow processing cannot be found easily or if situation C from above applies (long processing times due to a high overall load). Important Chapter 9 provides the basic checks for the hardware of the PI server. It should not only be used when analyzing a hardware bottleneck due to a high overall load and/or an insufficient sizing but should also be used after every change made to the configuration of PI to ensure that the hardware is able to handle the new situation. This is important because a classical PI system uses three runtime engines (IS, BPE, AFW). Tuning one engine for high throughput might have a direct impact on the others. With every tuning action applied, you have to be aware of the consequences on the other runtimes or the available hardware resources.

11

SAP NETWEAVER 7.1 AND HIGHER

3

PERFORMANCE CHECK GUIDE

DETERMINING THE BOTTLENECK

The total processing time in PI consists of the processing time in the Integration Server, processing time in the Business Process Engine (if ccBPM is used), as well as the processing time in the Adapter Framework (if adapters except IDoc, plain HTTP, or ABAP proxies are used). In the subsequent chapters you will find a detailed description of how to obtain these processing times. For reasons of completeness we will also have a look at the ABAP Proxy runtime and the available performance evaluations. 3.1

Integration Engine Processing Time

You can use an aggregated view of the performance (details about activation in Note 820622 - Standard jobs for XI performance monitoring) data for the Integration Engine as shown below: Procedure Log on to your Integration Engine and call transaction SXMB_IFR. In the browser window, follow the link to the Runtime Workbench. In there, click “Performance Monitoring”. Change the display to “Detailed Data Aggregated”, select ‘IS’ as the Data Source (or ‘PMI’ if you have configured the Process Monitoring Infrastructure’) and “XIIntegrationServer/” as the component. Then choose an appropriate time interval, for example, the last day. You have to enter the details of the specific interface you want to monitor here.  The interesting value is the “Processing Time [s]” in the fifth column. It is the sum of the processing time of the single steps in the Integration Server. For now you are not interested in the single steps that are listed on the right side since you are still trying to find out the part of the PI system where the most time is lost. Chapter 4.4 will work with the individual steps.

12

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

If you do not know which interface is affected yet, you first have to get an overview. Instead of navigating to Detailed Data Aggregated in the Runtime Workbench, choose Overview Aggregated. Use the button Display Options and check the options for sender component, receiver component, sender interface, and receiver interface. Check the processing times as described above.

3.2

Adapter Engine Processing Time

Procedure Log on to your Integration Server and call transaction SXMB_IFR. In the browser window follow the link to the Runtime Workbench. In there, click Message Monitoring, choose Adapter Engine from Database from the drop down lists, and press Display. Enter your interface details and “Start” the selection. Select one message using the radio button and press Details.  Calculate the difference between the start and end timestamp indicated on the screen.  Do the above calculation for the outbound (AFW  IS) as well as the inbound (IS  AFW) messages.  The audit log of successful message is no longer persisted in SAP NetWeaver PI 7.1 by default in order to minimize the load on the database. The audit log has a high impact on the database load since it usually consists of many rows that are inserted individually in the database. In newer releases therefore a cache was implemented that keeps the audit log information for a period of time only. Thus, no detailed information is available any longer from a historical point of view. Instead you should use Wily Introscope for historical analyses of the performance of successful messages.  The audit log can be persisted on the database to allow historical analysis of performance problems on the Java stack. To do so, change the parameter “messaging.auditLog.memoryCache” from true to false for service “XPI Service: Messaging System” using NetWeaver Administrator (NWA) as described in SAP Note 1314974. Note: Only do this temporarily if you have identified the bottleneck on the AFW. First try using Wily to determine the root cause of the problem.

13

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

 Due to above mentioned limitations Wily Introscope is the right tool for performance analysis of the Adapter Engine. Wily persists the most important steps like queuing, module processing and average response time in historical dashboards (aggregated data per intervals, for all server nodes of a system or all systems, with filters possible, etc).

14

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Especially for Java-Only runtime using Integrated Configuration Wily shows the complete processing time after the initial sender adapter steps (receiver determination, mapping and time in the backend call) as can be seen in the Message Processing Time dashboard below:

Using the “Show Minimum and Maximum” functionality deviations from the average processing time can be seen. In the example below a message processing step takes up to 22 seconds. The reason for this could either be a slow mapping or a delay in the call to the receiver system (in the example below Java IDoc adapter is used to transfer the data to ERP):

In case of high processing times in one of your steps further analysis might be required using thread dumps or the Java Profiler as outlined in the appendix section A.2 XPI inspector for troubleshooting and performance analysis. 3.2.1

Adapter Engine Performance monitor in PI 7.31 and higher

Starting from PI 7.31 SP4 there is a new performance monitor available for the Adapter Engine. More information on the activation of the performance monitor can be found at Note 1636215 – Performance Monitoring for the Adapter Engine. Similar to the ABAP monitor in the RWB the data is displayed in an aggregated format. On the PI start page use the link “Configuration and Monitoring Home”. Go to “Adapter Engine” and select “Performance Monitor”. The data is persisted on an hourly basis for the last 24 hours and on a daily basis for the last 7 days. The data is further grouped on interface level and per Java server node. With the information provided you can therefore see the minimum, maximum and average response time of an individual interface on a specific Java server node. All individual steps of the message processing like time spent in the Messaging System queues or Adapter modules are listed. In the example below you can 15

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

see that most of the time (5 seconds) is spent in a module called SimpleWaitModule. This would be the entry point for the further analysis.

Starting with 7.31 SP10 (7.40 SP05) this data can be exported to excel in two ways: Overview or Detailed (details in SAP Note 1886761). 3.3

Processing Time in the Business Process Engine

Procedure Log on to your integration server and call transaction SXMB_MONI. Adjust your selection in a way that you select one or more messages of the respective interface. Once the messages are listed, navigate to the Outbound column (or Inbound if your integration process is the sender) and click on PE. Alternatively, you can select the radio button for “Process View” on the entrance/selection screen of SXMB_MONI.  To judge the performance of an integration process, it is essential to understand the steps that are executed. A long-running integration process in itself is not critical because it can wait for a trigger event or can include a synchronous call to a backend. Therefore, it is important to understand the steps that are included before analyzing an individual workflow log. For example, the screenshot below shows a long waiting time after the mapping. This is perhaps due to an integrated wait step.  Calculate the time difference between the first and the last entry. Note that this time for the workflow does not include the duration of RFC calls, for example. To see this processing time, navigate to the “List with Technical Details” (second button from the left on the screenshot below or shift + F9). Repeat this step for several messages to get an overview about the most time consuming steps.  Please Note that the processing time mentioned above does not include the waiting time of the messages in the ccBPM queues. Therefore it is essential to monitor a queue backlog in ccBPM queues as discussed in section “Queuing and ccBPM (or increasing Parallelism for ccBPM Processes)”.

16

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Via transaction SWI2_DURA you can display an aggregated performance overview across all ccBPM processes running on your PI system. In the initial screen you have to choose “(Sub-) Workflow and the time range you would like to look at.

17

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

The results here allow you to compare the average performance of a process. It shows you the processing time based on barriers. The 50% barrier means that 50% of the messages were processed faster than the value specified. Furthermore you also see a comparison of the processing time to e.g. the day before. By adjusting the selection criteria of the transaction you can therefore get a good overview about the normal processing time of the Integration process and can judge if there is a general performance problem or just a temporary one. The number of process instances per integration process can be easily checked via transaction SWI2_FREQ. The data shown allows you to check the volume of the Integration Processes and allow you to judge if your performance problem is caused by an increase of volume which could cause a higher latency in the ccBPM queues. New in PI 7.3 and higher Starting from PI 7.3 a new monitoring for ccBPM processes is available. This monitor can be started from transaction SXMB_MONI_BPE  Integration Process Monitoring (also available in “Configuration and Monitoring Home” on PI start page). This is new browser based view that allows a simplified and aggregated view on the PI Integration Processes. On the initial screen you get an overview about all the Integration Processes executed in the selected time interval. Therefore you can immediately see the volume of each Integration Process.

From there you can navigate to the Integration process facing the performance issues and look at the individual process instances and the start and end time. Furthermore there is a direct entry point to see the PI messages that are assigned to this process.

18

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Choosing one special process instance ID you will see the time spend in each individual processing step within the process. In the example below you can see that most of the time is spend in the Wait Step:

19

SAP NETWEAVER 7.1 AND HIGHER

4

PERFORMANCE CHECK GUIDE

ANALYZING THE INTEGRATION ENGINE

If chapter Determining the Bottleneck has shown that the bottleneck is clearly in the Integration Engine, there are several transactions that can help you to analyze the reason for this. To understand why this selection of transactions helps to analyze the problem, it is important to know that the processing within the Integration Engine is done within the pipeline. The central pipeline of the Integration Server executes the following main steps:

Central Pipeline Steps

Description of Central Pipeline Steps

PLSRV_XML_VALIDATION_RQ_INB

XML Validation Inbound Channel Request

PLSRV_RECEIVER_DETERMINATION

Receiver Determination

PLSRV_INTERFACE_DETERMINATION

Interface Determination

PLSRV_RECEIVER_MESSAGE_SPLIT

Branching of Messages

PLSRV_MAPPING_REQUEST

Mapping

PLSRV_OUTBOUND_BINDING

Technical Routing

PLSRV_XML_VALIDATION_RQ_OUT

XML Validation Outbound Channel Request

PLSRV_CALL_ADAPTER

Transfer to Respective Adapter: IDoc, HTTP, Proxy, AFW

PLSRV_XML_VALIDATION_RS_INBXML

Validation Inbound Channel Response

PLSRV_MAPPING_RESPONSE

Response Message Mapping

PLSRV_XML_VALIDATION_RS_OUT

Validation Outbound Channel Response

The last three steps highlighted in bold are available in case of synchronous messages only and reflect the time spent in the mapping of the synchronous response message. The steps indicated in red are new in SAP NetWeaver PI 7.1 and higher and can be used to validate the payload of a PI message against an XML schema. These steps are optional and can be executed at different times in the PI pipeline processing, for example, before and after a mapping (as shown above). With PI 7.31 an additional option of using an external Virus Scan during PI message processing can be activated as described in the Online Help. The Virus Scan can be configured on multiple steps in the pipeline – e.g. after the mapping (Virus Scan Outbound Channel Request) when receiving a synchronous response (Virus Scan Inbound Channel Response) and after the response mapping (Virus Scan Outbound Channel Response). It is important to know is that these pipeline steps are executed based on qRFC Logical Unit of Work (LUW) by using dialog work processes.

20

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

With 7.11 and higher versions you can activate the Lean Message Search (LMS) to be able to search for payload content. This is described in more detail in chapter Long Processing Times for “LMS_EXTRACTION”. 4.1

Work Process Overview (SM50/SM66)

The work process overview is the central transaction to get an overview of the current processes running in the Integration Engine. For message processing PI is only using Dialog work processes (DIA WP). Therefore it is essential to ensure that enough DIA WPs available. The process overview is only a snapshot analysis. You therefore have to refresh the data multiple times to get a feeling for the dynamics of the processes. The most important questions to be asked are as follows:  How many work processes are being used on average? Use the CPU Time (clock symbol) to check that not all DIA WPs are used. In case all DIA WPs have a high CPU time this indicates a WP bottleneck.  Which users are using these processes? It depends on the usage of your PI system. All asynchronous messages are processed by the QIN scheduler who will start the message processing in DIA WPs. The user that is shown in SM66 will be the one that triggered the QIN scheduler. This can e.g. be a communication user for a connected backend (usually a copy of PIAPPLUSER) or the user used to send data from the Adapter Engine (per default PIAFUSER).  Activities related to the Business Process Engine (ccBPM) will be indicated by user WF-BATCH. If you see high WF-BATCH activity, take a look at chapter 5.1  Are there any long running processes and, if so, what are these processes doing with regard to the reports used and the tables accessed? o

As a rule of thumb for the initial configuration, the number of DIA work processes should be around six times the value of CPU in your PI system (rdisp/wp_no_dia = 6 to 8 times #CPUs cores based on SAP Note 1375656 - SAP NetWeaver PI System Parameters).

If you would like to get an overview for an extended period of time without actually refreshing the transaction at regular intervals, use report /SDF/MON and schedule the Daily Monitoring. It allows you to collect metrics such as CPU usage, amount of free Dialog work processes, and, most importantly, a snapshot of the work process activity as provided in transactions SM50 and SM66. The frequency for the data collection can be as low as 10 seconds and the data can be viewed after the configurable timeframe of the analysis. Note: /SDF/MON is only intended for analysis of performance issues and should only be scheduled for a limited period of time. If you have Solution Manager Diagnostics installed and all the agents connected to PI, you can also monitor the ABAP resource usage using Wily Introscope. As you can see in the dashboard below, you can use it to monitor the number of free work processes (prerequisite is that the SMD agent is running on the PI system). The major advantage of Wily Introscope is that this information is also available from the past and allows analysis after the problem has occurred in the system.

21

SAP NETWEAVER 7.1 AND HIGHER

4.2

PERFORMANCE CHECK GUIDE

qRFC Resources (SARFC)

Depending on the configuration of the RFC server group, not all dialog work processes can be used for qRFC processing in the Integration Server. As stated earlier all (ABAP based) asynchronous messages are processed using qRFC. Therefore tuning the RFC layer is one of the most important tasks to achieve good PI performance. In case you have a high volume of (usually runtime critical) synchronous scenarios you have to ensure that enough DIA WPs are kept free by the asynchronous interfaces running at the same time. Since this is a very difficult tuning exercise we usually recommend implementing runtime critical synchronous interfaces using Java only configuration (ICO) whenever possible. If this is not possible you have to ensure via SARFC tuning that enough work processes are kept free to process the synchronous messages. The current resource situation in the system can be monitored using transaction SARFC. Procedure First check which application servers can be used for qRFC inbound processing by checking the AS Group assigned in transaction SMQR. Call transaction SARFC and refresh several times since the values are only snapshot values.  Is the value for “Max. Resources” close to your overall number of dialog work processes? Max. Resources is a fixed value and describes how many DIA work processes may be used for RFC communication. If the value is too low your RFC parameters need tuning to provide the Integration Server with enough resource for the pipeline processing. The RFC parameters can be displayed by double-clicking on “Server Name”. o

A good starting point is to set the values as follows: Max. no. of logons = 90 Max. disp. of own logons = 90 Max. no. of WPs used = 90 Max. wait time = 5 Min. no. of free WP = 3-10 (depending on the size of the application server). These values are specified in SAP Note 1375656 – SAP NetWeaver PI System Parameters (this parameter must be increased if you plan to have synchronous interfaces). To avoid any bottleneck and blocking situation, free DIA WPs should always be available. The value should be adjusted based on the overall number of DIA WPs available at the system and the above requirements.

22

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Note: You have to set the parameters using the SAP instance profile. Otherwise the changes are lost after the server is restarted.  The field “Resources” shows the number of DIA WPs available for RFC processing. Is this value reasonably high, for example, above zero all the time? If the “Resources” value is zero, then the Integration Server cannot process XML messages because no dialog work process is free to execute the necessary qRFC. There are several conclusions that can be drawn from this observation as follows: 1) Depending on the CPU usage, (see Chapter “Monitoring CPU Capacity”) it might be necessary at this time to increase the number of dialog work processes in your system. If the CPU usage, however, is already very high, increasing the number of DIA will not solve the bottleneck. 2) Depending on the number of concurrent PI queues (see Chapter qRFC Queues (SMQ2)) it might be necessary to decrease the degree of parallelism in your Integration Server because the resources are obviously not sufficient to handle the load. The next section describes how to tune PI inbound and outbound queues.

The data provided in SARFC is also collected and shown in a graphical and historical way using Solution Manager Diagnostic and Wily Introscope (as can be seen in the screenshot below). This allows easy monitoring of the RFC resources across all available PI application servers.

4.3

Parallelization of PI qRFC Queues

The qRFC inbound queues on the ABAP stack are one of the most important areas in PI tuning. Therefore it is essential to understand the PI queuing concept. PI generally uses two types of queues for message processing – PI inbound and PI outbound queues. Both types are technical qRFC inbound queues and can therefore be monitored using SMQ2. PI inbound queues are named XBTI* (EO) or XBQI* (EOIO) and are shared between all interfaces running on PI by default. The PI outbound queues are named XBTO* (EO) and XBQO* (EOIO). The queue suffix (in red: XBTO0___0004) specifies the receiver business system. This way PI is using dedicated outbound queues for each receiver 23

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

system. All interfaces belonging to the same receiver business system are contained in the same outbound queue. Furthermore there are dedicated queues for prioritization of separation of large messages. To get an overview about the available queues use SXMB_ADM  Manage Queues. PI inbound and outbound queues execute different pipeline steps: PI Inbound Queues Pipeline Steps

Description of Pipeline Steps

PLSRV_XML_VALIDATION_RQ_INB

XML Validation Inbound Channel Request

PLSRV_RECEIVER_DETERMINATION

Receiver Determination

PLSRV_INTERFACE_DETERMINATION

Interface Determination

PLSRV_RECEIVER_MESSAGE_SPLIT

Branching of Messages

PI Outbound Queues Pipeline Steps

Description of Pipeline Steps

PLSRV_MAPPING_REQUEST

Mapping

PLSRV_OUTBOUND_BINDING

Technical Routing

PLSRV_XML_VALIDATION_RQ_OUT

XML Validation Outbound Channel Request

PLSRV_CALL_ADAPTER

Transfer to Respective Adapter: IDoc, HTTP, Proxy, AFW

Before the steps above are executed the messages are placed in a qRFC queue (SMQ2). The messages will wait till they are first in queue and till a free DIA work process is assigned to a queue. The wait time in the queue are recorded in the performance header in DB_ENTRY_QUEUING (PI inbound queue) and DB_SPLITTER_QUEUING (PI outbound queue). Most often you will see the long processing times there as discussed in chapter Long Processing Times for “DB_ENTRY_QUEUEING” and Long Processing Times for “DB_SPLITTER_QUEUEING”. Looking at the steps above it is clear that tuning the queues will have a direct impact on the connected PI components and also backend systems. For example, by increasing the number of parallel outbound queues more mappings will be executed in parallel which will in turn put a greater load on the Java stack or more messages will be forwarded in parallel to the backend system in case of a Proxy call. Thus, when tuning PI queues you must always keep the implications for the connected systems in mind. The tuning of PI ABAP queue parameters is done in transaction SXMB_ADM  Integration Engine Configuration and by selecting the category TUNING For productive usage we always recommend to use inbound and outbound queues (parameter EO_INBOUND_TO_OUTBOUND=1). Otherwise only inbound queues will be used, which are shared across

24

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

all interfaces. Hence a problem with one single backend system will affect all interfaces running on the system. The principle of less is sometimes more also applies for tuning the number of parallel PI queues. Increasing the number of parallel queues will result in more parallel queues with fewer entries per queue. In theory this should result in a lower latency if enough DIA WPs are available. But practically this is not true for high volume systems. The main reason for this is the overhead involved in the reloading of queues (see details below). Furthermore important tuning measures like PI packaging (see Chapter “PI Message Packaging”) aim to increase the throughput based on a higher number of messages in the queue. Thus, from a throughput perspective it is definitely advisable to configure fewer queues. If you have very high runtime requirements you should prioritize these interfaces and assign a different parallelism for high priority queues only. This can be done using parameters EO_INBOUND_PARALLEL_HIGH_PRIO and EO_OUTBOUND_PARALLEL_HIGH_PRIO as described in section “Message Prioritization on the ABAP Stack”. To tune the parallelism of inbound and outbound queues the relevant parameters are EO_INBOUND_PARALLEL and EO_OUTBOUND_PARALLEL have to be used. Below you can see a screenshot of SMQ2. You can see the PI inbound queues and outbound queues. Also ccBPM queues (XBQO$PE) are displayed and will be discussed in section 5.

Procedure Log on to your Integration Server, call transaction SMQ2, and execute. If you are running ABAP proxies on a separate client on the same system, enter ‘*’ for the client. Transaction SMQ2 provides snapshots only, and must therefore be refreshed several times to get viable information. o

The first value to check is the number of queues concurrently active over a period of time. Since each queue needs a dialog work process to be worked on, the number of concurrent queues must not be too high compared to the number of dialog work processes usable for RFC communication (see Chapter qRFC Resources (SARFC)). The optimum throughput can be achieved if the number of active queues is equal to or slightly higher than the number of dialog work processes (provided that the number of DIA work processes can be handled by the CPU of the system, see transaction ST06). In case many of the queues are EOIO queues (e.g. because the serialization is done on the material number) try to reduce the queues by following Chapter EOIO tuning.

o

The second value of importance is the number of entries per queue. Hit refresh a few times to check if the numbers increase/decrease/remain the same. An increase of entries for all queues, or a specific

25

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

queue, points to a bottleneck or a general problem on the system. The conclusion that can be drawn from this is not simple. Possible reasons have been found to include: 1) A slow step in a specific interface Bad processing time of a single message or a whole interface can be caused by expensive processing steps as for example, the mapping step or receiver determination. This can be confirmed by looking at the processing time for each step as shown in “Analyzing the runtime of pipeline steps”. 2) Backlog in Queues: Check if inbound or outbound queues face a backlog. A backlog in a queue is generally nothing critical since the queues ensure that the PI components as well as the backend systems are not overloaded. For instance batch triggered interfaces are usually causing high backlogs that get processed over time. Only in case the backlog prevents you to meet the business requirements for an interface this should be analyzed. If the backlog is caused by a high volume of messages arriving in a short period of time one solution to minimize the backlog is to increase the number of queues available for this interface (if sufficient hardware is available). If one interface is having a backlog in the outbound queues you could e.g. specify EO_OUTBOUND_PARALLEL with a sub parameter specifying your interface to increase the parallelism for this interface. But in general a backlog can also be caused by a long processing time in one of the pipeline steps as discussed in 1) or a performance problem with one of the connected components like PI Java stack or connected backend systems. Due to the steps performed in the outbound queues (mapping or call adapter) it is more likely that a backlog will be seen in the outbound queues. Inbound queue processing is only expensive if Content Based Routing or Extended Receiver Determination is used. To understand which step is taking long follow once more chapter “Analyzing the runtime of pipeline steps”. In addition to tuning of the number of inbound and outbound queues also “Message Prioritization on the ABAP Stack” and “PI Message Packaging” can help. The screenshot below shows such a backlog situation in Wily Introscope. Please Note that the information about the queues is only collected by Wily in 5 minutes intervals explaining the gaps between the measurement points. On the dashboard “Total qRFC Inbound Queue Entries” you can see that the number of messages in SMQ2 is increasing. The number of inbound queues does not increase at the same time. It also offers a dashboard for the number of entries by queue name and the time entries remain in SMQ2. With the information provided here, it is also possible to distinguish a blocking situation or a general resource problem after the problem is no longer present in the system.

26

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

3) Queues stay in status READY in SMQ2: To see the status of the queues, use the filter button in SMQ2 as shown below.

It is a normal situation to see many queues in READY status for a limited amount of time. The reason for this is the way the QIN scheduler reloads LUWs (see point 2 below). A waiting time of up to 5 seconds per queuing step is considered normal during normal load situation (even if there is no backlog in the queue and enough resources are available). If you observe a situation where many queues are in READY status for longer period of time the following situations can apply:  A resource bottleneck with regard to RFC resources. Confirm by following section qRFC Resources (SARFC).  To avoid overloading a system with RFC load, the QIN scheduler only reloads new queues after a certain amount of queues have finished processing. This can lead to long waiting times in the queues as explained in SAP Note 1115861 - Behaviour of Inbound Scheduler after Resource Bottleneck. Since PI is a non-user system, we recommend setting rfc/inb_sched_resource_threshold to 3 as described in SAP Note 1375656 – SAP NetWeaver PI System Parameters.  Long DB loading times for RFC tables. If using Oracle, ensure that SAP Note 1020260 Delivery of Oracle statistics is applied.  Additional overhead during the pipeline executions occurs due to PI internal mechanisms. Per default a lock is set during processing of each message. This is no longer necessary and should 27

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

be switched off by setting the parameter LOCK_MESSAGE of category RUNTIME to 0 as described in SAP Note 1058915. In addition during the processing of an individual message the message repeatedly calls an enqueue, which leads to a deterioration of the throughput. This can be avoided by setting parameter CACHE_ENQUEUE to 0 as described in SAP Note 1366904.  Sometimes a queue stays in READY status for a very long time while other queues are getting processed fine. After manual unlocking the queue is processed but then stops after some time again. A potential reason is described in Note 1500048 - SMQ2 inbound Queue remains in READY state. In a PI environment queues in status RUNNING for more than 30 minutes usually happen due to one individual step taking long (as discussed in step 1) or a problem on the infrastructure (e.g. memory problems on Java or HTTP 200 lost on network). After 30 minutes the QIN scheduler removes this queue from the scheduling and therefore it remains in READY. To solve such cases the root cause for the blocking has to be understood. 4) Queue in READY status due to queue blacklisting: If a queue is for a very long time in status RUNNING (e.g. due to a problem with a mapping or due to a very slow backend) it is possible that the QIN scheduler excludes this queue from queue scheduling. Hence the queue stays in status READY even though enough work processes are available. The “blacklisting” of queues takes place when the runtime of a queue exceeds the “Monitoring threshold” configured on this queue. Notes 1500048 - queues stay in ready (schedule_monitor) and Note 1745298 - SMQR : Inbound queue scheduler : identify excluded entries provide more input on this topic. The screenshot below shows a setting of 2400 seconds (40 minutes) for this parameter.

 Check if a queue stays in READY status for a long time while others are processing without any issue. Ensure Note 1745298 is implemented and check the system log for the following exception “QINEXCLUDE . If this is the case check why the long runtime occurs (see section Analyzing the runtime of PI pipeline steps) or increase the schedule_monitor in exceptional cases only (if e.g. the runtime of an individual processing step cannot be improved further). 5) An error situation is blocking one or more queues:  Click the Alarm Bell (Change View) push button once to see only queues with error status. Errors delay the queue processing within the Integration Server and may decrease the throughput if, for example, multiple retries occur. Navigate to the first entry of the queue and analyze the problem by forward navigation to the relevant message in SXMB_MONI.  Often you see EO queues in SYSFAIL or RETRY due to the problems during the processing of an individual message. This can be prevented by following the description of Chapter Prevent blocking of EO queues.

28

SAP NETWEAVER 7.1 AND HIGHER

4.4

PERFORMANCE CHECK GUIDE

Analyzing the runtime of PI pipeline steps

The duration of the pipeline steps is the ultimate answer to long processing times in the Integration Engine since it describes exactly how much time was spent at which point. The recommended way to retrieve the duration of the pipeline steps is the RWB and will be described below. Advanced users may use the Performance Header of the SOAP message using transaction SXMB_MONI, but the timestamps are not easy to read. If you still prefer the latter method, here is the explanation: 20110409092656.165 must be read as yyyymmddhh(min)(min)(sec)(sec).(microseconds), that is, the timestamp corresponds to April 9, 2011, at 09:26:56 and 165 microseconds. Please note that these timestamps in Performance Header of SXI_MONITOR transaction are stored and displayed in UTC time, therefore conversion to system time must be done when analyzing them. In case PI message packaging is configured the performance header will always reflect the processing time per package. Hence a duration of 50 seconds does not mean a single message took 50 seconds but eventually the package contained 100 messages, so that every message took 0,5 ms. More details about this can be found in section PI Message Packaging. Procedure Log on to your Integration Server and call transaction SXMB_IFR. In the browser window follow the link to the Runtime Workbench. In there, click Performance Monitoring. Change the display to Detailed Data Aggregated and choose an appropriate time interval, for example, the last day. For this selection you have to enter the details of the specific interface you want to monitor.  You will now see in the lower–part of the screen a split of the processing time into single steps. Check the time difference between the steps. Does any step take longer than expected? In the example in the screenshot below, the DB_ENTRY_QUEUEING starts after 0.032 seconds and ends after 0.243 seconds which means it took 0.211 seconds (211ms).

29

SAP NETWEAVER 7.1 AND HIGHER



PERFORMANCE CHECK GUIDE

Compare the processing times for the single steps for different measurements as outlined in Chapter Pipeline Steps (SXMB_MONI or RWB). For example, is a single step only long if many messages are processed or if a single message is processed? This helps you to decide if the problem is a general design problem (single message has long processing step) or if it is related to the message volume (only for a high number of messages this process step has large values).

Each step has different follow-up actions that are described next.

4.4.1

Long Processing Times for “PLSRV_RECEIVER_ DETERMINATION” / PLSRV_INTERFACE_DETERMINATION

Steps that can take a long time in inbound processing are the Receiver and Interface Determination. In these steps, the receiver system and interface is calculated. Normally this is very fast but PI offers the possibility of enhanced receiver determinations. In these cases the calculation is based on the payload of a message. There are different implementation options: o

Content-Based Routing (CBR): CBR allows defining XPath expressions that are evaluated during runtime. These expressions can be combined by logical operators, for example, to check the value for multiple fields. The processing time of such a request directly correlates to the number and complexity of the conditions defined.

 No tuning options exist for the system in regard to CBR. The performance of this step can only be changed by reducing the number of rules by changing the design of the interfaces. Another option is to use an extended receiver determination (executing a mapping program) since this is faster than CBR using many rules. o

Mapping to determine Receivers: A standard PI mapping (ABAP, Java, XSLT) can also be used to determine the receiver. If you observe high runtimes in such a receiver determination, follow the steps outlined in the next section since the same mapping runtime is used.

4.4.2

Long Processing Times for “PLSRV_MAPPING_REQUEST”

Before analyzing the mapping, you must understand which runtime is used. Mappings can be implemented in ABAP, as graphical mappings in the Enterprise Service Builder, as self-developed Java mappings or XSLT mappings. One interface can also be configured to use a sequence of mappings executed sequentially. In such a case analysis is more difficult because it is not clear which mapping in the sequence is taking a long time. Normal debugging and tracing tools (transaction ST12) can be used for mappings executed in ABAP. Any type of mapping can be tested from ABAP using transaction SXI_MAPPING_TEST. Knowing sender/receiver party/service/interface/namespace and source message payload it is possible to check the target message (after mapping execution) and detailed trace output (similar to contents of trace of the message in SXMB_MONI with TRACE_LEVEL = 3). It can also be used for debugging at runtime by using the standard debugging functionality. For ABAP based XSLT mappings it is also possible via transaction XSLT_TOOL to test, trace and debug XSLT transformations on the ABAP stack. In general the mapping response time is heavily influenced by the complexity of the mapping and the message size. Therefore, to analyze a performance problem in the mapping environment you should compare the mapping runtime during the time of the problem with values reported several days earlier to get a better understanding.

30

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

If no AAE is used, the starting point of a mapping is the Integration Engine that will send the request by RFC destination AI_RUNTIME_JCOSERVER to the gateway. There it will be picked up by a registered server program. The registered server program belongs to the J2EE Engine. The request will be forwarded to the J2EE Engine by a JCo call and then executed by the Java runtime. When the mapping has been executed, the result is sent back to the ABAP pipeline. There are therefore multiple places to check when trying to determine why the mapping step took so long. Before analyzing the mapping runtime of the PI system, check if only one interface is affected or if you face a long mapping runtime for different interfaces. To do so, check the mapping runtime of messages being processed at the same time in the system. The best tool for such an analysis is Wily Introscope which offers a dashboard for all mappings being executed at a given time. Each line in the dashboard represents one mapping and shows the average response time and the number of invocations. In the screenshot below you can see that many different mapping steps have required around 500 seconds for processing. Comparing the data during the incident with the data from the day before will allow you to judge if this might be a problem of the underlying J2EE engine as described in section J2EE Engine Bottleneck.

If there is only one mapping that faces performance problems, there would be just one line sticking out in the Wily graphs. If you face a general problem that affects different interfaces, you can choose a longer timeframe that allows you to compare the processing times in a different time period and verify if it is only a “temporary” issue – this would, for example, indicate an overload of the mapping runtime. If you have found out that only one interface is affected, then it is very unlikely to be a system problem but rather a problem in the implementation of the mapping of that specific interface.

31

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

 Check the message size of the mapping in the runtime header using SXMB_MONI. Verify if the message size is larger than usual (which would explain the longer runtime).

 There can also be one or many lookups to a remote system (using RFC or JDBC) causing the long processing time of a mapping. Together with the application, you then have to check if the connection to the backend is working properly. Such a mapping with lookups can be analyzed using Wily Transaction Trace as explained in the appendix section Wily Transaction Trace. If not one but several interfaces are affected, a potential system bottleneck occurs and this is described in the following. o

Not enough resources (registered server programs) available. That could either be the case if too many mapping requests are issued at the same time or if one J2EE server node is down and has not registered any server programs.

 To check if there were too many mapping requests for the available registered server programs, compare the number of outbound queues that are concurrently active with the number of registered server programs. The number of outbound queues that are concurrently active can be monitored with transaction SMQ2 and counting queues with the names XBTO* & XBQO*/XB2O*, XBTA*, and XBQA**/XB2A* (high priority) XBTZ* and XBQZ*/XB2Z* (low priority), and XBTM* (large messages). In addition to this you have to take into account mapping steps that are executed by synchronous messages or in a ccBPM process. Mappings executed by a ccBPM process are not queued, but the ccBPM process calls the mapping program directly using tRFC. The number of registered server programs can be determined with transaction SMGW  Goto  Logged On Clients and filtering by the program (“TP Name”) AI_RUNTIME_. In general we recommend configuring 20 parallel mapping connections per server node.

 If the problem occurred in the past, it is more difficult to determine the number of queues that are concurrently active. One way is to use transaction SXMB_MONI and specify every possible outbound queue (field “Queue ID”) in the advanced selection criteria (Note: wildcard search not available). The timeframe can be restricted in the upper-part of the advanced selection criteria. Advanced users could use transaction SE16 to search directly in table SXMSPMAST, using the field “QUEUEINT”. The other option is to use the Wily “qRFC Inbound Queue Detail” dashboard as described in qRFC Queue Monitoring (SMQ2).  To check if the J2EE is not available, or if the server program is not registered for a different reason, use transaction SM59 to test the RFC destination AI_RUNTIME_JCOSERVER. If the problem occurred in the past, search the gateway trace (transaction SMGW  GoTo  Trace  Gateway  Display File) and the RFC developer traces dev_rfc* files available in the work directory for the string “AI_RUNTIME”. In addition, check the std_server.out in the work directory for restarts of the J2EE Engine.

32

SAP NETWEAVER 7.1 AND HIGHER

o

PERFORMANCE CHECK GUIDE

Depending on the checks above, there are two options to resolve the bottleneck. Of course, if the J2EE was not available, the reason for this has to be found and prevented in the future. The two options for tuning are: 1) the number of outbound queues that are concurrently active and 2) the number of mapping connections from the J2EE server to the ABAP gateway.

The increase of mapping connections is only recommended for strong J2EE servers since each mapping thread needs resources like CPU and Heap memory. Too many mapping connections mean too many mappings being executed on the J2EE Engine. In turn, this might reduce the performance of other parts of the PI system, for example, the pipeline processing in the Integration Engine or the processing within the adapters of the AFW, or can even cause out-of-memory errors on the J2EE engine.  Add additional J2EE server nodes to balance the parallel requests. Each J2EE server node will register new destinations at the gateway and will therefore take a part of the mapping load.  Another option is to reduce the number of outbound queues that are concurrently active to solve the bottleneck. This will certainly result in higher backlogs in the queue and is therefore only applicable for less runtime critical interfaces. This option can be used if the backlog is caused by a master data interface with high volume, but no critical processing time requirements. In such a case you can only lower the number of parallel queues for this interface by using a sub-parameter of EO_OUTBOUND_PARALLEL in transaction SXMB_ADM. By doing so, you slow down one interface to ensure other (more critical) interfaces have sufficient resources available. o

4.4.3

If multiple application servers configured on the PI system ensure that all the server nodes are connected to their local gateway (local bundle option), as described in the guide How To Scale PI. Long Processing Times for “PLSRV_CALL_ADAPTER”

Call adapter is the last step executed in the PI pipeline and forwards the message to the next component along the message flow. In most cases (local Adapter Engine, IDoc, or BPE) this is a local call. The IDoc adapter only puts the message on the tRFC layer (SM58) of the PI system so that the actual transfer of the IDoc is not included in the performance header. For ABAP Proxies, plain HTTP or calls to a central or decentral Adapter Engine an HTTP call is made so that network time can have an influence here. Looking at the processing time, we have to distinguish asynchronous (EO and EOIO) and synchronous (BE) interfaces. In asynchronous interfaces, the call adapter step includes the transfer of the message by the network and the initial steps on the receiver side (in most cases this just means persisting the message on the receiver’s database). A long duration can therefore have two reasons: o

Network latency: For large messages of several MB in particular, the network latency is an important factor (especially if the receiving ABAP proxy system is located on another continent, for example). Network latency has to be checked in case of a long call adapter step. In case HTTP load balancers are used they are considered as part of the network.

o

Insufficient resources on receiver side: Enough resources must be available at the receiver side to ensure quick processing of a message. For instance, enough ICM threads and dialog work processes must be available in the case of an ABAP proxy system. Therefore, the analysis of long call adapter steps always has to include the relevant backend system.

For synchronous messages (request/response behavior) the call adapter step also includes the processing time on the backend to generate the response message. Therefore the call adapter for synchronous messages includes the time of the transfer of the request message, the calculation of the corresponding response message at the receiver side, and the transfer back to PI. Therefore, the processing time to process a request at the receiving target system for synchronous messages must always be analyzed to find the most costly processing steps.

33

SAP NETWEAVER 7.1 AND HIGHER

4.4.4

PERFORMANCE CHECK GUIDE

Long Processing Times for “DB_ENTRY_QUEUEING”

The value for DB_ENTRY_QUEUEING describes the time that a message has spent waiting in a PI inbound queue before processing started. In case of errors the time also includes the wait time for restart of the LUW in the queue. The inbound queues (XBTI*, XBT1*, XBT9*, XBTL* for EO messages and XBQI*/XB2I*, XBQ1*/XB21*,, XBQ9**/XB29* for EOIO messages) process the pipeline steps for the Receiver Determination, the Interface Determination, and the Message Split (and optionally XML inbound validation). Thus, if the step DB_ENTRY_QUEUEING has a high value, the inbound queues have to be monitored using transactions SMQ2 and SARFC. The reasons are similar as those outlined in chapter qRFC Resources (SARFC) and qRFC Queues (SMQ2). o

Not enough resources (DIA work processes for RFC communication) available.

o

A backlog in the queues caused by one of the following reasons:  The number of parallel inbound queues is too low to handle the incoming messages in parallel. Note that a simple increase of the parameter EO_INBOUND_PARALLEL might not always be the solution (as enough work processes also have to be available to process the queues in parallel). The number of work processes is in turn restricted by the number of CPUs that are available for your system. If you would like to separate critical from non-critical sender systems, define a dedicated set of inbound queues by specifying a sender ID as a sub parameter of EO_INBOUND_PARALLEL. For example, by doing so you can separate master data interfaces from business critical interfaces. Please note that the separation on inbound queues only works on Business System level and not on interface level.  The inbound queues were blocked by the first LUW. Use transaction SMQ2 to check if that is still the case. To identify if such a situation occurred in the past, use the Wily Introscope dashboard “ABAP System qRFC Inbound Detail“. It is generally not easy to find the one message that is blocking the queue. It might be achieved by checking RFC traces or by searching for a single message with a long pipeline processing step. For EO queues there is no business reason to block the queues in case of a single message error. To prevent this follow the description in section Prevent blocking of EO queues.  In case you have a different runtime of the messages in the queues (e.g. due to complex extended Receiver Determinations) the number of messages per queue might be uneven distributed between the queues. In PI 7.3 a new balancing option is available as discussed in Avoid uneven backlogs with queue balancing.  Problems when loading the LUW by the QIN scheduler as described in qRFC Queues (SMQ2).

Note: If your system uses the parameter EO_INBOUND_TO_OUTBOUND = 0, you must also read chapter 4.4.5 for analyzing the reasons. EO_INBOUND_TO_OUTBOUND only determines whether inbound queues (value ‘0’) or inbound and outbound queues (value ‘1’) are used for the pipeline processing in the integration server. Check the value with transaction SXMB_ADM  Integration Engine Configuration  Specific Configuration. The default value is ‘1’, meaning the usage of inbound and outbound queues (the recommended behavior). 4.4.5

Long Processing Times for “DB_SPLITTER_QUEUEING”

The value for DB_SPLITTER_QUEUEING describes the time that a message has spent waiting in a PI outbound queue until a work process was assigned. In case of errors the time also includes the wait time for the restart of the LUW in the queue. The outbound queues (XBTO*, XBTA*, XBTZ*, XBTM* for EO messages and XBQO*/XB2O*, XBQA*/XB2A*, XBQZ*/XB2Z* for EOIO messages) process the pipeline steps for the mapping, outbound binding and call adapter.

34

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Thus, if the step DB_SPLITTER_QUEUEING has a high value, the outbound queues have to be monitored using transactions SMQ2 and SARFC. The reasons are similar as outlined in chapters qRFC Resources (SARFC) and qRFC Queues (SMQ2) or described in the section above for PI inbound queues. o

Not enough resources (DIA work processes for RFC communication) available.

o

A backlog in the queues caused by one of the following reasons:  The number of parallel outbound queues is too low to handle the outgoing messages in parallel. Note that simply increasing the parameter EO_OUTBOUND_ PARALLEL might not always be the solution as enough work processes have to be available to process the queues in parallel. The number of work processes in turn is limited by the number of CPUs available for the system. Also, the parameter EO_OUTBOUND_PARALLEL is used differently than the parameter EO_INBOUND_ PARALLEL because it determines the degree of parallelism for each receiver system. Thus, it is possible to increase the number of parallel outbound queues for specific receivers whereas other receivers are handled with a lower degree of parallelism. Note that not every receiver backend is able to handle a high degree of parallelism so that in case of an ABAP Proxy system it might not indicate a problem on PI in case you observe a high value for DB_SPLITTER_QUEUING.  The outbound queues were blocked by the first LUW. To identify if such a situation occurred in the past, use the Wily Introscope dashboard “ABAP System qRFC Inbound Detail“. In general it is not easy to find the one message that is blocking the queue. It might be achieved by checking RFC traces or by searching for a single message with a long pipeline processing step. For EO queues there is no business reason to block the queues in case of a single message error. To prevent this follow the description in section Prevent blocking of EO queues.  Since the outbound queues include a processing step that is bound to take longer than the other steps (that is, the mapping and the call adapter step) this might mean a long waiting time for all subsequent messages. As an example, assume that a mapping takes 1 second and that 100 messages are waiting in a queue. The first message gets executed at once, the second message has to wait about 1 second (a bit more since more steps are executed than just the mapping, but this should be ignored at the moment), the third takes 3 seconds, and so on. The 100th message has to wait 100 seconds until it is processed, that is, the value for DB_SPLITTER_QUEUING is as high as 100. For a given outbound queue you would therefore see an increase for the DB_SPLITTER_QUEUING value over time. If you experience this situation, proceed with Chapter 4.4.2.  In case you have a different runtime of the messages in the queues (e.g. due to different message size) the number of messages per queue might be uneven distributed between the queues. In PI 7.3 a new balancing option is available as discussed in Avoid uneven backlogs with queue balancing.

o

Problems when loading the LUW by the QIN scheduler as described in qRFC Queues (SMQ2).

Note: The number of parallel outbound queues is also connected with the ability of the receiving system to process a specific amount of messages for each time unit.  In section Adapter Parallelism, default restrictions of the specific adapters of the J2EE Engine are explained. Some adapters (JDBC, File, Mail) can process only one message for each server node at a time (this can be changed). It does not make sense for such adapters to have too many parallel qRFC queues since this will only move the message backlog from ABAP to Java.  For messages that are directed to the IDoc outbound adapter, the value for EO_OUTBOUND_PARALLEL is connected to the MAXCONN value of the corresponding outbound destination. You can define the maximum number of connections using transaction SMQS – the default value is 10 parallel connections. If applicable, SAP highly recommends that you use IDoc packaging in the outbound IDoc adapter to reduce the load during the transmission of tRFC calls from the PI system to the receiving system.  For ABAP proxy systems, the number of outbound queues directly determines the number of messages sent in parallel to the receiving system. Thus, you have to ensure that the resources 35

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

available on PI are aligned with those on the sending/receiving ABAP proxy system. 4.4.6

Long Processing Times for “LMS_EXTRACTION”

Lean Message Search can be configured for newer PI releases as described in the Online Help. After applying Note 1761133 - PI runtime: Enhancement of performance measurement an additional header for LMS will be written to the performance header. The header could look like this indicating that around 25 seconds were spent in the LMS analysis:

When using trace level two additional timestamps are written to provide details about this overall runtime:  LMS_EXTRACTION_GET_VALUES This timestamp describes the evaluation of the message according to user-defined filter criteria.  LMS_EXTRACTION_ADJUST_VALUES This timestamp describes the post-processing of the previously found values in the BAdI (as of PI 7.30). The runtime of the LMS heavily depends on the number of payload attributes to be extracted. Also the payload size and the complexity of the XPath expression has a direct impact on the performance of the LMS. In general the number of elements to be indexed should be kept at a minimum and very deep and complex XPath expressions should be avoided. If you want to minimize the impact of LMS on the PI message processing time you can define the extraction method to use an external job. In that case the messages will be indexed after processing only and it will therefore have no performance impact during runtime. Of course this method imposes a delay in the indexing (based on the frequency of job SXMS_EXTRACT_MESSAGES) so that newly processed messages cannot be searched using LMS. If this delay is acceptable for the monitoring responsible the messages should be indexed using an external job.

4.4.7

Other step performed in the ABAP pipeline

As discussed earlier there are additional steps in the ABAP pipeline that can be executed based on the configuration of your scenario. Per default these steps will not be activated and should therefore not consume any time. All these steps are traced in the performance header of the message. Below you can see the details for: -

XML validation:

36

SAP NETWEAVER 7.1 AND HIGHER

-

PERFORMANCE CHECK GUIDE

Virus Scan:

In case one of these steps is taking long you have to check the configuration of your scenario. In the example of the Virus scan the problem might be related to the external virus scanner used and tuning has to happen on that side. 4.5

PI Message Packaging for Integration Engine

PI message packaging was introduced with SAP NetWeaver PI 7.0 SP13 and is per default activated in PI 7.1 and higher. Message packaging in BPE is independent of message packaging in PI (see chapter 5.6), but they can be used together. To improve performance and message throughput, asynchronous messages can be assembled to packages and processed as one entity. For this the qRFC scheduler was enabled to process a set of messages (a package) in one LUW. Thus, instead of sending one message to the different pipeline steps (for example, mapping/routing), a package of messages will be sent that will reduce the number of context switches that is required. Furthermore, access to databases is more efficient since requests can be bundled in one database operation. Depending on the number and size of messages in the queue, this procedure improves performance considerably. In return, message packaging can increase the runtime for individual messages (latency) due to the delay in the packaging process. Message packaging is only applicable for asynchronous scenarios (QoS EO and EOIO). Due to the potential latency of individual messages, packaging is not suitable for interfaces with very high runtime requirements. In general packaging has the highest benefit for interfaces with high volume and small message size. The performance improvement achieved directly relates to the number of the messages bundled in each package. Message packaging must not solely be used in the PI system. Tests have shown that the performance improvement significantly increases if message packaging is configured end-to-end - that is, from the sending system using PI to the receiving system. Message packaging is mainly applicable for application systems connected to PI by ABAP proxy. From the runtime perspective, no major changes are introduced when activating packaging. Messages remain individual entities in regards to persistence and monitoring. Additional transactions are introduced allowing the monitoring of the packaging process. Messages are also treated individually for error handling. If an error occurs in a message processed in a package, then the package will be disassembled and all messages will be processed as single messages. Of course, in a case where many errors occur (for example, due to interface design) this will reduce the benefits of message packaging. SAP systems connected to PI via ABAP Proxy can also make use of the message packaging to reduce the HTTP calls necessary to transmit messages between the systems. Other sending and receiving applications in particular will not see any changes because they send and receive individual messages and receive individual commits. The PI message packaging can be adapted to meet your specific needs. In general, the packaging is determined by three parameters: 37

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

1) Message count: Maximum number of messages in a package (default 100) 2) Maximum package size: Sum of all messages in kilobytes (default 1 MB) 3) Delay time: Time to wait before the queue is processed if the number of messages does not reach the message count (default 0 meaning no waiting time) These values can be adjusted using transactions SXMS_BCONF and SXMS_BCM. To better use the performance improvements offered by packaging, you could, for example, define a specific packaging for interfaces with very small messages to allow up to 1,000 messages for each package. Another option could be to increase the waiting time (only if latency is not critical) to create bigger packages. See SAP Note 1037176 - XI Runtime: Message Packaging for more details about the necessary prerequisites and configuration of message packaging. More information is also available at http://help.sap.com/: SAP NetWeaver  SAP NetWeaver PI/Mobile/IdM 7.1  SAP NetWeaver Process Integration 7.1 Including Enhancement Package 1  SAP NetWeaver Process Integration Library  Function-Oriented View  Process Integration  Integration Engine  Message Packaging. 4.6

Prevent blocking of EO queues

In general messages for interfaces using Quality of Service Exactly Once are independent of each other. In case of an error in one message there is no business reason to stop the processing of other messages. But exactly this happens in case an EO queue goes into error due to an error in the processing of a single message. The queue will then be automatically retried in configurable intervals. This retry will cause a delay of all other messages in the queue which cannot be processed due to the error of the first message in the queue. To avoid this, a new error handling was introduced in SAP Notes 1298448 - XI runtime: No automatic retry for EO message processing (set EO_RETRY_AUTOMATIC = 0 and schedule restart job RSXMB_RESTART_MESSAGES) and SAP Note 1393039 - XI runtime: Dump stops XI queue. After applying this Note EO queues will not go in SYSFAIL status any longer. These Notes are recommended for all PI systems and are per default activated on PI 7.3 systems. By specifying a receiver ID as sub parameter this behavior can be configured for individual interfaces only. In PI 7.3 you can also configure the number of messages that would go into error before the whole queue would be stopped. For permanent errors (e.g. due to a problem with the receiving backend) the queue would go into SYSFAIL so that not too many messages are going into error. This also reduces the number of alerts for Message Based Alerting. To define the threshold the parameter EO_RETRY_AUT_COUNT can be set. The default value is 0 and indicates that the number of messages is not restricted. 4.7

Avoid uneven backlogs with queue balancing

From PI 7.3 on a new mechanism is available to address the distribution of messages in queues. Per default in all PI versions the messages are getting assigned to the different queues randomly. In case of different runtimes of LUWs caused by e.g. different message sizes or different runtimes in mappings this can cause an uneven distribution of messages within the different queues. This can increase the latency of the messages waiting at the end of that queue. To avoid such an unequal distribution the system checks the queue length before putting a new message in the queue. Hence it tries to achieve an equal balancing during inbound processing. A queue which has a higher backlog will get less new messages assigned. Queues with fewer entries will get more messages assigned. Therefore it is different than the old BALANCING parameter of category TUNING) which was used 38

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

to rebalance messages already assigned to a queue. The assignment of messages to queues is shown in the diagram below:

To activate the new queue balancing you have to set the parameters EO_QUEUE_BALANCING_READ and EO_QUEUE_BALANCING_SELECT of category TUNING in SXMB_ADM. If the value of the parameter EO_QUEUE_BALANCING_READ is 0 (default value) then the messages are distributed randomly across the queues that are available. If however EO_QUEUE_BALANCING_READ is set to a value n greater than zero, then on average the current fill level of the queue is determined after every nth message and stored in the shared memory of the application server. This data is used as the basis for determining the queue (see description of parameter EO_QUEUE_BALANCING_SELECT) relevant for balancing. The parameter EO_QUEUE_BALANCING_SELECT specifies the relative fill levels of the queues in percent. Relative here means in relation to the maximum filled queue. If there are queues with a lower fill level than defined here then only these are taken into consideration for distribution. If all queues have a higher fill level then all queues are taken into consideration. Please note that determining the fill level takes place using database access and therefore impacts system performance. The value of EO_QUEUE_BALANCING_READ should for this reason be oriented towards the message throughput and specific requirements or even distribution. For higher volume system a higher value should be chosen. Let’s look at an example. In case you set the parameter EO_QUEUE_BALANCING_READ to 1000 after every 1000th incoming message the queue distribution will be checked. This requires a database access and can therefore cause a performance impact. The fill level for each queue will then be written to shared memory. In our example we configured EO_QUEUE_BALANCING_SELECT to 20. At a given point in time you have three outbound queues on the system. XBTO__A contains 500 entries, XBTO__B 150 and XBTO__C contains 50 messages. Therefore XBTO__B has a fill level of 30% and XBTO__C of 10%. That means the only queue relevant for balancing is XBTO__C which will get more messages assigned. Based on this example you can see that it is important to find the correct values. As a general guideline to minimize performance impact, you should set EO_QUEUE_BALANCING_READ to a rather large value. The correct value of EO_QUEUE_BALANCING_SELECT depends on the criticality of a delay caused by a backlog.

39

SAP NETWEAVER 7.1 AND HIGHER

4.8

PERFORMANCE CHECK GUIDE

Reduce the number of parallel EOIO queues

As discussed in Chapter Parallelization of PI qRFC Queues it is important to keep the number of queues limited. As a rule of thumb the number of active queues should be equal to the number of available work processes in the system. It is especially crucial to not have a lot of queues containing only one message since this will cause a high overhead on the QIN scheduler due to very frequent reloading of the queues. Especially for EOIO queues we often see a high number of parallel queues containing only one message. The reason for this is for example that the serialization has to be done on document number to e.g. ensure that updates to the same material are not transferred out of sequence. Typically during a batch triggered load of master data this can result in hundreds of EOIO queues in SMQ2 just containing only one or two messages. This is very bad from a performance point of view and will cause significant performance degradation for all other interfaces running at the same time. To overcome this situation the overall number of EOIO queues (for all interfaces) can be limited as of PI 7.1 as described in Change Number of EOIO Queues. The maximum number of queues can be set in SXMB_ADM  Configure Filter for Queue Prioritization  Number of EOIO queues.

During runtime a new set of queues will be used with the name XB2* as can be seen below for the outbound queues.

These queues will be shared between all EOIO interfaces for the given interface priority and by setting the number of queues the parallelization for all EOIO interfaces is limited. Thus more messages will be using the same EOIO queue and therefore PI message packaging will work better and also the reloading of the queues by the QIN scheduler will show much better performance. In case of errors the messages will be removed from the XB2* queues and will be moved to the standard XBQ* queues. All other messages for the same serialization context will be moved to the XBQ* queue directly to ensure the serialization is maintained. This means that in case of an error the shared EOIO queues will not be blocked and messages for other serialization contexts will be not delayed.

40

SAP NETWEAVER 7.1 AND HIGHER

4.9

PERFORMANCE CHECK GUIDE

Tuning the ABAP IDoc Adapter

Very often the IDoc adapter deals with very high message volume and tuning of this is very essential. 4.9.1

ABAP basis tuning

As stated above, the IDoc adapter uses the tRFC layer to transfer messages from PI to the receiver system. The IDoc adapter will only put the message in SM58 and from there the standard tRFC layer forwards the LUW. You therefore have to ensure that sufficient resources (mainly DIA WPs) are available for processing tRFC requests. Wily Introscope also offers a dashboard (shown below) for the tRFC layer which shows the tRFC entries and their different statuses. With these dashboards you are also able to identify history backlogs on tRFC.

In order to control the resources used when sending the IDocs from sender system to PI or from PI to the receiver backend, you can also think about registering the RFC destinations in SMQS in PI as known from standard ALE tuning. By changing the Max. Conn. or the Max. Runtime values in SMQS you can limit or increase the number of DIA WPs used by the tRFC layer to send the IDocs for a given destination. This will mitigate the risk of system overload on sender and receiver side.

41

SAP NETWEAVER 7.1 AND HIGHER

4.9.2

PERFORMANCE CHECK GUIDE

Packaging on sender and receiver side

One option for improving the performance of messages processed with the IDoc adapter is to use IDoc packaging. The receiver IDoc adapter in PI (for sending messages from PI to the receiving application system) already had the option for packaging IDocs in previous releases. For this, messages processed in the IDoc adapter are bundled together before being sent out to the receiving system. Thus, only one tRFC call is required to send multiple IDocs. The message packaging that is being discussed in section PI Message Packaging uses a similar approach and therefore replaces the “old” packaging of IDocs in the receiver IDoc adapter. Therefore we highly recommend configuring Message Packaging since this helps transferring data for the IDoc adapter as well as the ABAP Proxy. For the sender IDoc adapter there was previously the option to activate packaging in the partner profile of the sending system in case IDocs are not send immediately but in batch. By specifying the “Maximum Number of IDocs” in the report RSEOUT00 multiple IDocs are bundled in one tRFC calls. At the PI side, these packages will be disassembled by the IDoc adapter and the messages will be processed individually. Therefore this will help mainly to reduce the tRFC resources involved in transferring the data between the systems but not within PI. This behavior was changed with SAP NetWeaver PI 7.11 and higher: If the sender backend sends an IDoc package to PI, a new option has been introduced which allows the processing of IDocs as a package on PI as well. You can specify an IDoc package size on the sender IDoc Communication Channel. The necessary configuration is described in detail in the following SDN blog - IDoc Package Processing Using Sender IDoc Adapter in PI EhP1. A significant increase in throughput was recorded for high volume interfaces with small IDocs. 4.9.3

Configuration of IDoc posting on receiver side

As described in SAP Note 1828282 - IDoc adapter: long runtime XI -> IDoc-> ALE also the way the IDocs are posted on the receiver ERP backend will directly influence the processing of the LUW on the PI side. In general two options exist for IDoc posting:

 Trigger immediately: The IDoc will be posted directly when it is received. For this a free dialog workprocess will be required.

 Trigger by background program: If this option is chosen the IDoc is persisted on the ERP side and the posting of the application data will happen asynchronously using report RBDAPP01 via a background job.

42

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

When “trigger immediately” is configured the sender system (PI) has to wait till the IDoc is posted. While the workprocess will roll out you will see the LUW in “Transaction Executing” during this time. For the IDOC_AAE, the RFC connection and the java threads will be kept busy until the IDOC is posted in the backend, therefore leading to backlogs on the IDOC_AAE.

The coding for posting the application data can be very complex and therefore this operation can take a long time consuming unnecessary resources also on the sender side. With background processing of the IDocs the sender and receiver systems are decoupled from each other. Due to this we generally recommend to use “trigger immediately” in exceptional cases where the data has to be posted without any delay. In all other cases like high volume replication of master data background processing of IDocs should be chosen. With Note 1793313 and 1855518 a third option for posting IDocs on the ERP side was introduced. This option is using bgRFC on the ERP side to queue the IDocs. As soon as the configured resources for the bgRFC destination are available the IDocs will be posted. By doing so the IDocs will be posted based on resource availability on the receiver system and no additional background jobs will be required. Furthermore storing the LUWs in bgRFC will ensure that the sender and receiver systems are decoupled. This option therefore guarantees that the IDocs will be posted as soon as possible based on the available resources on the receiver system without the requirement to schedule many background jobs. This new option is therefore the recommended way of IDoc posting in systems that fulfill the necessary requirements. 43

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

4.10 Message Prioritization on the ABAP Stack A performance problem is often caused by an overload of the system, for example, due to a high volume master interface. If business critical interfaces with a maximum response time are running at the same time then they might face delays due to a lack of resources. Furthermore, as described in qRFC Queues (SMQ2), messages of different interfaces use the same inbound queue by default which means that critical and lesscritical messages are mixed up in the same queue. We highly recommend using message prioritization on the ABAP stack to prevent such delays for critical interfaces (for Java, see Interface Prioritization in the Messaging System ). For more information about PI prioritization, see http://help.sap.com and navigate to SAP NetWeaver  SAP NetWeaver PI/Mobile/IdM 7.1  SAP NetWeaver Process Integration 7.1 Including Enhancement Package 1  SAP NetWeaver Process Integration Library  Function-Oriented View  Process Integration  Integration Engine  Prioritized Message Processing. To balance the available resources further between the configured interfaces, you can also configure a different parallelization level for queues with different priorities. Details for this can be found in SAP Note 1333028 – Different Parallelization for HP Queues.

44

SAP NETWEAVER 7.1 AND HIGHER

5

PERFORMANCE CHECK GUIDE

ANALYZING THE BUSINESS PROCESS ENGINE (BPE)

This chapter helps to analyze performance problems if Chapter Determining the Bottleneck has shown that the BPE is the slowest step for a message during its time in the PI system. Of course, this type of runtime engine is very susceptible to inefficient design of the implemented integration process. Information about best practices for designing BPE processes can be found in the documentation Making Correct Use of Integration Processes. In general, the memory and resource consumption is higher than for a simple pipeline processing in the Integration Engine. As outlined in the document linked above, every message that is sent to BPE and every message that is sent from BPE is duplicated. In addition, work items are created for the integration process itself as well as for every step. More database space is therefore required than previously expected and more CPU time is needed for the additional steps. 5.1

Work Process Overview

The Business Process Engine executes the steps (work items) of an integration process using tRFC calls to the RFC destination WORKFLOW_LOCAL_. To check if there are enough resources available to process the work items, call transaction SM50 (on each application server) while one of the integration processes is running. o

Look for the user “WF-BATCH” and follow its activities by refreshing the transaction

o

Are there always dialog work processes available?

The most important point to keep in mind here is the concurrent activity of the Business Process Engine (for ccBPM processes) and the Integration Engine (for pipeline processing). Both runtime environments need dialog work processes. Thus, performance problems in one of the engines cannot be solved by restricting the other. Rather, an appropriate balance between the two runtimes has to be found. The activity of the WF-BATCH user is not restricted by default. It is highly recommended that you register the RFC destination WORKFLOW_LOCAL_ with transaction SMQS and assign a suitable amount of maximum connections. Otherwise, ccBPM activity may block the pipeline processing of the Integration Server and reduce the throughput of PI drastically. SAP Note 888279 - Regulating/distributing the Workflow Load gives more details. In addition, it might help to use specific servers (dialog instances) to carry out a given workflow or to use load balancing. Of course, both options only apply for larger PI systems that consist of at least a central instance and one dialog instance. 5.2

Duration of Integration Process Steps

As described in section Processing Time in the Business Process Engine with PI 7.3 there is a new monitor available in “Configuration and Monitoring Home” on PI start page. It shows the duration of individual steps of an integration process in a very easy way and is the now tool to analyze performance related issues on the ccBPM.

45

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

For releases prior than PI 7.3 check in the workflow log for long-running steps as described in the previous chapter. Use the “List with Technical Details” to do so. Since the design of integration processes varies significantly, there is no general rule for the duration of a step or for specific sequences of steps. Instead, you have to get a feeling for your implemented integration processes.  Did a specific process step decrease in performance over a period of time?  Does one specific process step stick out with regard to the other steps of the same integration process?  Do you observe long durations for a transformation step (“mapping”)?  Do you observe long durations of synchronous send/receive steps in the integration process which correlates, for example, to a slow response from a connected remote system? Use the SAP Notes database to learn about recommendations and hints. SAP Note 857530 - BPE: Composite Note Regarding Performance acts as a central note for all performance notes and might be used as the entry point. Once you are able to answer the above questions, there are several paths to follow: o

Two common reasons should be taken into consideration for a performance decrease over time: Firstly, the load on PI might have increased over time as well. Check the hardware resources (chapter 9) and carefully monitor the work process availability (chapter 5.1). Alternatively, the underlying databases might slow down the integration processes. Use chapter 5.4 to check this possibility.

o

If it is a specific process step that sticks out, the process design itself has to be reviewed. The SAP Notes mentioned above might help here.

o

If the mapping step takes a long time, it is worth having a look at chapter 4.4.2 since the BPE and the IE use the same resources (mapping connections from the ABAP gateway to the J2EE server) to execute their mapping.

o

If there is a long processing time for a synchronous send/receive step, check the connected backend system for any performance issues.

46

SAP NETWEAVER 7.1 AND HIGHER

5.3

PERFORMANCE CHECK GUIDE

Advanced Analysis: Load Created by the Business Process Engine (ST03N)

Since the Business Process Engine runs on the ABAP part of the PI Integration Server, you can use the statistical data written by the dialog work processes to analyze a performance problem. Procedure Log in to your Integration Server and call transaction ST03N. Switch to “Expert Mode” and choose the appropriate timeframe (this could be the “Last Minute’s Load” from the Detailed Analysis menu or a specific day or week from the Workload menu). In the Analysis View, navigate to User and Settlement Statistics and then to User Profile. The user you are interested in is the WF-BATCH user who does all ccBPM-related work.

 Compare the total CPU Time (in seconds) to the time frame (in seconds) of the analysis. In the above example, an analysis time frame of 1 hour has been chosen which corresponds to 3600 seconds. To be exact, these 3600 seconds have to be multiplied by the number of available CPUs. Out of these 3600*#CPU seconds, the WF-BATCH user used up 36 seconds. o

The above value gives you a good idea of how much load the Business Process Engine creates on your PI Integration Server. By comparison with the other users (as well as the CPU utilization from transaction ST06), you can determine if the Integration Server is able to handle this load with the available number of CPUs or not. This does not help you to optimize the load, but it does provide support when deciding how to distribute the workload of the BPE and the workload of the Integration Server (see Chapter 5.1).

 Experienced ABAP administrators should also analyze the database time, wait time, and compare these values. 5.4

Database Reorganization

The database is accessed many times during the processing of an Integration Process. This is not usually connected with performance problems, but if a specific database table is large then statements may take longer than for small database tables.  Use transaction ST05 to collect a SQL trace while the integration process in question is being executed. Restrict the trace to user WF-BATCH. When the integration process is finished, stop the trace and view it. Look for high values in the duration column, especially for SELECT statements.  Use transaction SE16 to determine the number of entries for the typical workflow tables: SWW_WI2OBJ and SWFRXI*. There are many more database tables used by BPE, but if the number of entries is high for the above tables, they are high for the rest as well. See SAP Note 872388 Troubleshooting Archiving and Deletion in XI and the links given in this note if you find high amounts of entries. 47

SAP NETWEAVER 7.1 AND HIGHER

5.5

PERFORMANCE CHECK GUIDE

Queuing and ccBPM (or increasing Parallelism for ccBPM Processes)

Until recently, ccBPM processes accepted messages from only one queue (one queue XBQO$PEWS* for each workflow). If there was a high message throughput for this workflow then a high backlog occurred for these queues. A couple of enhancements have therefore been implemented to improve scalability if the ccBPM runtime: o

Inbound processing without buffering

o

Use of one configurable queue that can be prioritized or assigned to a dedicated server

o

Use of multiple inbound queues (queue name will be XBPE_WS*)

o

Configure transactional behavior of ccBPM process steps to reduce number of persistency steps

Details about the configuration and tuning of the ccBPM runtime are described at: https://www.sdn.sap.com/irj/sdn/howtoguides  SAP Netweaver 7.0  End-to-End Process Integration: o

How to Configure Inbound Processing in ccBPM Part I: Delivery Mode

o

How to Configure Inbound Processing in ccBPM Part II: Queue Assignment

o

How to Configure ccBPM Runtime Part III: Transactional Behavior of an Integration Process

Procedure  Use transaction SMQ2 to check if the queues of type XBQO$PE* show a high backlog. o

5.6

Based on this information, verify if the ccBPM process is suitable to run with multiple queues. This is especially the case for collect patterns realized in BPE. If you can use multiple queues, configure the workflow based on the above guides. Message Packaging in BPE

Inbound processing takes up the most amount of processing time in many scenarios within BPE. Message packaging in BPE helps to improve performance by delivering multiple messages to BPE process instances in one transaction. This can lead to an increased throughput, which means that the number of messages that can be processed in a specific amount of time can increase significantly. Message packaging can also increase the runtime for individual messages (latency) due to the delay in the packaging process. The sending of packages can be triggered when the packages exceed a certain number of messages, a specific package size (in kB), or a maximum waiting time. The extent of the performance improvement that can be obtained depends on the type of scenario. Scenarios with the following prerequisites are generally most suitable: o

Many messages are received for each process instance

o

Messages that are sent to a process instance arrive together in a short period of time

o

Generally high load on the process type

o

Messages do not have to be delivered immediately

For example, "collect scenarios" are particularly suitable for message packaging. The higher the number of messages is in a package, the higher the potential performance improvements will be. Tests that have been performed have shown a high potential for throughput improvements up to factor 4.7 in tested Collect Scenarios depending on the packaging size that has been configured. For more details about the functionality of Message Packaging in BPE and its configuration, see SAP Note 1058623 - BPE: Message Packaging.

48

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Message packaging in BPE is independent of message packaging in PI (see Chapter 4.5), but they can be used together.

49

SAP NETWEAVER 7.1 AND HIGHER

6

PERFORMANCE CHECK GUIDE

ANALYZING THE (ADVANCED) ADAPTER ENGINE

This chapter describes further checks and possible reasons for bottlenecks in the (Advanced) Adapter Engine. Although this is not the place to describe the architecture of the Adapter Framework in detail, the following aspects are important to know for the analysis of a performance problem on the AFW: o

The AFW contains the Messaging System that is responsible for the persistence and queuing of messages. Within the flow of a message, it is placed in between the sender adapter (for example, a file adapter that polls from a folder) and the Integration Server as well as between the Integration Server and the receiver adapter (for example, a JMS adapter that sends a message out to an external system).

o

The Messaging System uses 4 queues for each adapter type and these are responsible for receiving and sending messages. To separate the message transfer, each adapter (for example, JMS, SOAP, JDBC, and so on) has its own set of queues. One queue for receiving synchronous messages (Request queue), one queue for sending synchronous messages (Call queue), one queue for receiving asynchronous messages (Receive queue), and one queue for sending asynchronous messages (Send queue). The name of the queue always states the adapter type and the queue name (for example, JMS_http://sap.com/xi/XI/SystemReceive for the JMS receiver asynchronous queues).

o

A configurable number of consumer threads are assigned on each queue. These threads work on the queue in parallel, meaning that the Messaging Queues are not strictly FirstIn-FirstOut. Five threads are assigned by default to each queue and thus five messages can be processed in parallel. But, as will be discussed later, the parallel execution of requests to the backend depends heavily on the adapter being used since not every adapter can work in parallel by default.

o

SAP NetWeaver PI 7.1 introduced a new queue: The dispatcher queue. The dispatcher queue is used to realize prioritization on the AAE (see chapter Prioritization in the Messaging System). Every message in the PI system is first assigned to the dispatcher queue before it is assigned to the adapter-specific queues.

Viewed from the perspective of a message that enters the PI system using a J2EE Engine adapter (for example File) and leaves the PI system using a J2EE Engine adapter (for example, JDBC) asynchronously, the steps are as follows: 1) Enter the sender adapter 2) Put into the dispatcher queue of the messaging system 3) Forwarded to the File send queue of the messaging system (based on Interface priority) 4) Retrieved from the send queue by a consumer thread 5) Sent from the messaging system to the pipeline of the ABAP Integration Engine (if no Integrated Configuration (AAE) is used 6) Processed by the pipeline steps of the Integration Engine 7) Sent to the messaging system by the Integration Engine 8) Put into the dispatcher queue of the messaging system 9) Forwarded to the JDBC receive queue of the messaging system (based on Interface priority) 10)

Retrieved from the receive queue (based on the maxReceivers)

11)

Sent to the receiver adapter

12)

Sent to the final receiver.

50

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

All the analysis is based on the information provided in the Audit Log. With SAP NetWeaver PI 7.1 the audit log is not persisted for successful messages in the database by default to avoid performance overhead. Therefore, the audit log is only available in the cache for a limited period of time (based on the overall message volume). More details can be found in chapter Persistence of Audit Log information in PI 7.10 and higher. As described in chapter Adapter Engine Performance monitor in PI 7.31 and higher a new performance monitor is available from PI 7.31 SP4. With this monitor you can display the processing time per interface in a given interval and identify the processing steps in which a lot of time is spend:

With 7.31 SP10 and 7.4 SP5 a download functionality for the performance monitor is provided as described in SAP Note 1886761. 6.1

Adapter Performance Problem

Compared to a bottleneck in the Messaging System it is relatively easy to detect a performance problem/bottleneck in the adapter. Since the timestamps for inbound and outbound messages are written at different points in time, the procedure is described separately for sender and receiver adapters. Note: It is not generally possible for all sender and receiver adapters to work in parallel (discussed in section Adapter Parallelism). 6.1.1

Adapter Parallelism

As mentioned before, not all adapters are able to handle requests in parallel. Thus, increasing the number of threads working on a queue in the messaging system will not solve a performance problem/bottleneck. There are 3 strategies to work around these restrictions: 1) Create additional communication channels with a different name and adjust the respective sender/receiver agreements to use them in parallel. 2) Add a second server node that will automatically have the same adapters and communication channel running as the first server node. This does not work for polling sender adapters (File, JDBC, or Mail) since the adapter framework scheduler assigns only one server node to a polling communication channel. 3) Install and use a non-central Adapter Framework for performance-critical interfaces to achieve better separation of interfaces.

51

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Some of the most frequently-used adapters and the possible options are discussed below: o

Polling Adapters (JDBC, Mail, File): At the sender side these adapters use an Adapter Framework Scheduler which assigns a server node that does the polling at the specified interval. Thus, only one server node in the J2EE cluster does the polling and no parallelization can be achieved. Therefore scaling via additional server node is not possible. More details will be presented in section Adapter Framework Scheduler. For example, since the channels are doing the same SELECT statement on the database or pick up files with the same file name, parallel processing will only result in locking problems. To increase the throughput for such interfaces, the polling interval has to be reduced to avoid a big backlog of data that is to be polled. If the volume is still too high, you should think about creating a second interface, for example, the new interface would poll the data from a different directory or database table to avoid locking. At the receiver side, the adapters work sequentially on each server node by default. For example, only one UPDATE statement can be executed for JDBC for each Communication Channel (independent of the number of consumer threads configured in the Messaging System) and all the other messages for the same Communication Channel will wait until the first one is finished. This is done to avoid blocking situations on the remote database. On the other hand, this can cause blocking situations for whole adapters as discussed in section Avoid Blocking Caused by Single Slow/Hanging Receiver Interface. To allow better throughput for these adapters you can configure the degree of parallelism at the receiver side. In the Processing tab of the Communication Channel in the field “Maximum Concurrency”, enter the number of messages to be processed in parallel by the receiver channel. For example, if you enter the value 2, then two messages are processed in parallel on one J2EE server node. The parallel execution of these statements at database level of course depends on the nature of the statements and the isolation level defined on the database. If all statements update the same database record database locking will occur and no parallelization can be achieved.

o

JMS Adapter: The JMS adapter is per default using a push mechanism on the PI sender side. This means the data is pushed by the sending MQ provider. Per default every Communication channel has one JMS connection established per J2EE server node. On each connection it processes one message after the other so that the processing is sequential. Since there is one connection per server node scaling via additional server nodes is an option. With Note 1502046 - JMS Adapter message pull mechanism the JMS protocol can be changed to use a pull mechanism. By doing so you can specify a polling interval in the PI Communication Channel and PI will be the initiator of the communication. Also here the Adapter Framework Scheduler is used which implies sequential processing. But in contrary to JDBC or File sender channels the JMS polling sender channel will allow parallel processing on all server nodes of the J2EE cluster. Therefore scaling via additional J2EE server nodes is an option. The JMS receiver side supports parallel operation out of the box. Only some small parts during message processing (pure sending of the message to the JMS provider) are synchronized. Therefore to enable parallel processing on the JMS receiver side no actions are necessary.

o

SOAP Adapter: The SOAP adapter is able to process requests in parallel. The SOAP sender side has in general no limitations in the number of requests it can execute in parallel. The limiting factor here is the FCAThreads available to process the incoming HTTP calls. This is discussed in more detail in chapter Tuning SOAP sender adapter. On the receiver side the parallelism depends on the number of the threads defined in the messaging system and the ability of the receiving system to cope with the load.

52

SAP NETWEAVER 7.1 AND HIGHER

o

PERFORMANCE CHECK GUIDE

RFC Adapter: The RFC adapter offers parameters to adjust the degree of parallelism by defining the number of initial and maximum connections to be used. The initial threads are allocated from the Application thread pool directly and are therefore not available for any other tasks. Therefore the number of initial thread should be kept minimal. To avoid bottlenecks during the peak time the maximum connections can be used. A bottleneck is indicated by the following exception in the Audit Log: com.sap.aii.af.ra.ms.api.DeliveryException: error while processing message to remote system:com.sap.aii.af.rfc.core.client.RfcClientException: resource error: could not get a client from JCO.Pool: com.sap.mw.jco.JCO$Exception: (106) JCO_ERROR_RESOURCE: Connection pool RfcClient … is exhausted. The current pool size limit (max connections) is 1 connections. Also this should be done carefully since these threads will be taken from the J2EE application thread pool. Therefore a very high value there can cause a bottleneck on the J2EE engine and therefore cause a major instability of the system. As per RFC adapter online help the maximum number of connections are restricted to 50.

o

IDoc_AAE Adapter: The IDoc adapter on Java (IDoc_AAE) was introduced with PI 7.3. On the sender side the parallelization depends on the configuration mode chosen. In Manual Mode the adapter works sequential per server node. For channels in “Default Mode” it depends on the configuration of the inbound Resource Adapter (RA). Via the parameter MaxReaderThreadCount of the inbound RA you can configure how many threads are globally available for all IDoc adapters running in Default Mode. Hence, this determines the overall parallelization of the IDoc sender adapter per Java server node. Currently the recommended maximum number of threads is 10. The receiver side of the Java IDoc adapter is working parallel per default.

The table below gives a summary of the parallelism for the different adapter types.

53

SAP NETWEAVER 7.1 AND HIGHER

6.1.2

PERFORMANCE CHECK GUIDE

Sender Adapter

The first step(s) to be taken for a sender adapter depend on the type of adapter itself. Afterwards, however, the message flow is always the same. First the message is processed in the Module Processor and afterwards put into the queue of the Messaging System. If the message is synchronous then it is the call queue, if the message is asynchronous that it is the send queue. The final action of the Adapter Framework is then to send the message from the messaging system to the Integration Server (for Non ICO interfaces). Only the first entries of the audit log are of interest to establish if the sender adapter has a performance problem: From the first entry to the entry “Message successfully put into the queue”. These steps logically belong to the sender adapter and the Module Processor while the subsequent steps belong to the Messaging System. The sender adapter uses one J2EE Engine application thread to execute its work. This is generally the same thread for all steps involved.  Select the “Details” of a message in the AFW. The first entry of the audit log will be the beginning of the activity of the sender adapter. The end of the adapter’s activity is marked by the entry “The application sent the message asynchronously using connection *. Returning to application”.

6.1.2.1 Tuning SOAP sender adapter As described above in general there is no limitation in the parallelization of the SOAP sender adapter itself. The incoming SOAP requests are limited by the available FCA Threads for HTTP processing. The FCA Threads are discussed in more detail in FCA Server Threads. Per default all interfaces using the SOAP adapter are using the same set of FCA Threads since they are all using the same URL http://:/MessageServlet . In case one interface is facing a very high load or slow backend connections this can cause a blockage of the available FCA Threads and could cause a heavy impact on the processing time of other interfaces. To avoid this SAP Note 1903431 allows the usage of different URLs for specific interfaces. This way you could isolate interfaces with a very high volume on a dedicated entry point using its own set of FCA Threads. In case of high load for this interface the other SOAP sender interfaces would not be affected by a shortage of FCA Threads.

54

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

To use this new feature you can specify on the sender system the following two entry points: o

/MessageServletInternal

o

/MessageServletExternal

More information about the URL of the SOAP adapter can be found here. 6.1.3

Receiver Adapter

For a receiver adapter it is just the other way around when compared to the sender adapter. First the message is received from the Integration Server and then put into a queue of the Messaging System, this time either the request or the receive queue for asynchronous and the request queue for synchronous messages. The last step is then the activity of the receiver adapter itself.  Select the “Details” of a message in the AFW. The adapter activity starts after the entry “The message status set to DLNG.” The message will then be forwarded to the entry point of the AFW “Delivering to channel: ”, enters the Module Processor “MP: Entering Module Processor” and ends with “The message was successfully delivered to the application using connection *”.

 Depending on the type of the adapter, you now have to investigate why the receiver adapter needs a high amount of processing time. This could be a complicated operating system command, a bad network connection, or a slow backend system among many other reasons.  Here, too, custom modules can be the reason for a prolonged processing time. Check Performance of Module Processing for more details.  Note: From 7.1 the audit log is not persisted any longer for performance reasons as described in Persistence of Audit Log information in PI 7.10 and higher.

55

SAP NETWEAVER 7.1 AND HIGHER

6.1.4

PERFORMANCE CHECK GUIDE

IDoc_AAE adapter tuning

A comprehensive and up-to date description about the tuning of the IDoc adapter is summarized in Note 1641541 - IDoc_AAE Adapter performance tuning. Please refer to this Note for details. Similar to all other adapters running on the Adapter Engine also the IDOC_AAE adapter is using the Messaging System queues and the explanations of the next chapter Messaging System Bottleneck are also valid for the IDOC_AAE adapter. Note that the IDOC_AAE adapter only supports async. Scenarios – so it is only using the “Send” queue in Java-only scenarios and the “Send” and “Receive” queue in dual-stack scenarios.” For the IDoc_AAE sender you can configure in addition bulk support as indicated in the screenshot below. By doing so all messages received by PI in one RFC call from ERP will be processed as one message. This is similar to the mechanism of the ABAP IDoc adapter described in the chapter Packaging on sender and receiver side except with the difference that the number of messages in one bulk cannot be further limited.

It is also essential to ensure that the size of the IDocs or IDoc packages for sender or receiver IDocs is not getting too large to avoid any negative impact on the Java heap memory and Garbage Collection. In the case of the IDoc_AAE adapter, the XML size of the IDoc is less critical; it is the number of segments in an IDoc or the overall number of segments in all IDocs of an IDoc package that will influence the memory allocation during message processing. Based on the current implementation per IDoc segment around 5 KB of memory during processing is required. A package of 5 IDocs with 2000 segments for each IDoc would consume roughly 500 MB during processing. In such cases it is important to e.g. lower the package size and/or use large message queues as outlined in chapter Large message queues on PI Adapter Engine. 6.1.5

Packaging for Proxy (SOAP in XI 3.0 protocol) and Java IDoc adapter

Due to message packaging in the Integration Engine (see PI Message Packaging for Integration Engine) the ABAP based IDoc and Proxy adapter was always using packaging when sending asynchronous messages to the receiver system. With PI 7.31 packaging was also introduced for Java based IDoc and Proxy (SOAP in XI 3.0 protocol) adapter. The aim of packaging is to achieve similar benefits to packaging already existing in ABAP pipeline: o

Less calls to backend consuming less parallel Dialog Work Processes on the receiver

o

Less overhead due to context switches / authentication when calling the backend

Especially for high volume interfaces packaging should be enabled to avoid overloading of the backend systems and impact on users operating on the system.

56

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Important In the ABAP stack the packaging is active per default (in 7.1 and higher) as described in chapter PI Message Packaging for Integration Engine. Experience shows that packaging can have a very high impact on the DIA Work process usage on the ERP receiver side. Especially for migration projects from dual-stack PI to AEX/PO it is important to evaluate packaging to avoid any negative impact on the receiving ECC due to missing packaging.

The packaging on Java works in general different then on ABAP. While in ABAP the aggregation is done on the individual qRFC queue level (which always contains messages to one receiver system only) this is not possible on the Adapter Engine since messages for all receivers are located in the same queue. Therefore packages are built outside of the Adapter Engine queues. The messages to be packaged will be forwarded to a bulk handler which waits till either the time for the package is exceeded or the number of messages or data size is reached. After this the message is send by a bulk thread to the receiving backend system. Packaging on Java is always done for a server node individually. If you have a high number of server nodes packaging will work less efficiently due to the load balancing of messages across the available server nodes. When enabling message packaging the message status will stay in status “Delivering” throughout all steps described above. In the audit log you can see the time it spends in packaging. The audit log shown below shows e.g. that the message was almost waiting one minute before the package was built and 9 IDocs were sent with this package:

Packaging can be enabled globally by setting the following parameters in Messaging System service as described in Note 1913972: o

"messaging.system.msgcollector.enabled": Enable packaging or disable it globally

o

"messaging.system.msgcollector.maxMemPerBulk": The maximum size of a bulk message

o

"messaging.system.msgcollector.bulkTimeout"; The wait time per bulk - default 60 seconds

o

"messaging.system.msgcollector.maxMemTotal": Maximum memory which can be used by message collector.

o

"messaging.system.msgcollector.poolSize": Number of parallel threads used to send out packages to the backend. This value has to be tuned based on the general performance of the receiver system, the network latency between the systems or the configuration of the IDoc posting on the ERP sender side (described in chapter IDoc posting on receiver side). Therefore tuning of these threads might be very important. These threads can unfortunately not be monitored yet with any PI tool or Wily 57

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Introscope. To verify that not all packaging threads are in use you can use the Thread monitoring in the SAP Management Console and check for threads named “BULK_EXECUTOR” as shown below:

If you would like to adapt the packaging for specific communication channel this can also be done using the configuration options in the Integration Directory:

While in the SOAP adapter in XI 3.0 protocol you can define three different packaging KPIs this is currently not possible for the IDoc adapter. There you can only specify the package size based on number of messages:

6.1.6

Adapter Framework Scheduler

For polling adapters like File, JDBC and JMS the Adapter Framework Scheduler determines on which server node the polling takes place. Per default a File or JDBC sender only works on one server node and the Adapter Framework (AFW) Scheduler determines on which one the job is scheduled. The Adapter Framework's scheduler had several performance and reliability issues in a cluster environment, especially in systems having many Java server nodes and many polling communication channels. Based on this and the symptoms described in Note 1355715 - AF Scheduler to avoid using cluster communication a new scheduler was released. We highly recommend using the new version of the AFW scheduler. To activate the new scheduler you have to set the parameter scheduler.relocMode of the Adapter framework

58

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

service property to some negative value. For instance by setting the value to -15 after every 15th polling interval a rebalancing of the channel within the J2EE cluster might happen. This can allow for better balancing of the incoming load across the available server nodes if the files are coming in on regular intervals. The balancing is achieved by doing a servlet call. Based on the HTTP load balancing the channel might then be dispatched to another server node. To avoid the balancing overhead this value should not be set too low. A value between -10 and -15 should be acceptable in most cases. For very short polling intervals (e.g. every second) even lower values (e.g. -50) can be configured. In case many files are put at one given time on the directory (by e.g. using a batch process) all these files will be processed by one polling process only. Hence a proper load balancing cannot be achieved by the AFW scheduler. In such a case the only option is to use write the files with different names or different directories so that you can configure multiple sender channels to pick up the files. Starting from PI 7.31 you can monitor the Adapter Framework Scheduler in pimon  Monitoring  Background Job Processing Monitor  Adapter Framework Scheduler Jobs. There you can check on which server node the Communication Channel is polling (Status=”Active”). Also you can see the time the channel was polling the last time and will poll next time.

You cannot influence the server node on which the channel is polling. This is determined via the AFW scheduler. You can only influence the frequency after which a channel can potentially move to another server node by tuning the parameter relocMode as outlined above. Prior to PI 7.31 you have to use the following URL to monitor the Adapter Framework Scheduler: http://server:port/AdapterFramework/scheduler/scheduler.jsp.

Above you see an example of a File sender channel. The type value determines if the channel runs only on one server node (type = L (local)) or on all server nodes (Type = G (global)). The time in the brackets [*,60000,*] determines the polling interval in ms – in this case 60 seconds. The Status column shows the status of the channel: o

“ON”: Currently polling

o

“on”: Currently waiting for the next polling

59

SAP NETWEAVER 7.1 AND HIGHER

o

6.2

PERFORMANCE CHECK GUIDE

“off”: not scheduled any longer (e.g. channel deactivated or scheduled on another server node)

Messaging System Bottleneck

As described earlier, the Messaging System is responsible for persisting and queuing messages that come from any sender adapter and go to the Integration Server pipeline or Java only interfaces (ICO), or that come from the Integration Server pipeline and go to any receiver adapter. The time difference between receiving the message from the one side and delivering it to the other side is the value that can be used to analyze bottlenecks in the Messaging System. The following chapters describe how to determine the time difference for both directions. A bottleneck (or rather a “fake” bottleneck) in the Messaging System can usually be closely connected with a performance problem in the receiver adapter. You must therefore make sure you execute the check described in chapter 6.1.3. If the receiver adapter needs a lot of time to process messages, then of course the messages get queued in the messaging system and remain there for a long time since the adapter is not ready to process the next one yet. It looks like the messaging system is not fast enough, but actually the receiver adapter is the limiting factor.  Execute the checks described in chapter 6.2.1 and 6.2.2. to get an overview of the situation, you should do this for multiple messages.  Decide if a specific receiver adapter is causing long waiting times in the messaging system by executing the check described in chapter 6.1.3. An additional hint at a receiver adapter problem is if only messages to specific receivers show a long wait time in the queue.  Use the queue monitoring in the Runtime Workbench (RWB) to analyze how many threads are active for each of the queue types and how large the queue size currently is. To do so, choose Component Monitoring  Adapter Engine. Press the Engine Status button and choose tab “Additional Data”. This will open the page below showing the number of messages in a queue, the threads currently working, and the maximum available threads. Note: This view only shows the Messaging System of one J2EE server node. To get the whole picture you have to check all the server nodes that are available in your system via the drop down list box.

60

SAP NETWEAVER 7.1 AND HIGHER

o

PERFORMANCE CHECK GUIDE

Since checking all the server nodes with this browser frontend can be very tedious, it is better to use Wily Introscope since it allows easy graphical analysis of the queues and thread usage of the messaging system across Java server nodes.

The starting point in Wily Introscope is usually the PI triage showing all the backlogs in the queue. In the screenshot below a backlog in the dispatcher queue (for 7.1 and higher only) can be seen:

The dispatcher queue is not the problem here since it will forward messages to the adapter queue if any free threads are available on the adapter specific consumer threads. The analysis must start in the PI inbound queues. By using the navigation button in the upper right corner of the inbound queue size you can directly jump to a more detailed view where you can see that the file adapter was causing the backlog:

61

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

To see the consumer thread usage, you can then follow the link to the file adapter. In the screenshot below you can see that all consumer threads were exceeded during the time of the backlog. Thus, increasing the number of consumer threads could be a solution if the resulting delay is not acceptable for business:

o

If it is obvious that the messaging system does not have enough resources as in the case above, increase the number of consumers for the queue in question. Use the results of the previous check to decide which adapter queues need more consumers. The adapter specific queues in the messaging system have to be configured in the NWA using service “XPI Service: AF Core” and property "messaging.connectionDefinition". The default values for the sending and receiving consumer threads are set as follows: (name=global, messageListener=localejbs/AFWListener, exceptionListener= localejbs/AFWListener,, pollInterval=60000, pollAttempts=60, Send.maxConsumers=5, Recv.maxConsumers=5, Call.maxConsumers=5, Rqst.maxConsumers=5) To set individual values for a specific adapter type, you have to add a new property set to the default set with the name of the respective adapter type, for example: (name=JMS_http://sap.com/xi/XI/System, messageListener=localejbs/AFWListener, exceptionListener=localejbs/AFWListener, pollInterval=60000, pollAttempts=60, Send.maxConsumers=7, Recv.maxConsumers=7, Call.maxConsumers=7, Rqst.maxConsumers=7) The name of the adapter specific queue is based on the pattern _ , for example RFC_http://sap.com/xi/XI/System. Note that you must not change parameters such as pollInterval and pollAttempts. For more details, see SAP Note 791655 - Documentation of the XI Messaging System Service Properties.

o

Not all adapters use the parameter above. Some special adapters like CIDX, RNIF, or Java Proxy can be changed by using the service “XPI Service: Messaging System” and property messaging.connectionParams by adding the relevant lines for the adapter in question as described above.

o

A performance problem on the Messaging System can also be caused by a high logging as described in chapter Logging / Staging on the AAE (PI 7.3 and higher).

o

Important: Make sure that your system is able to handle the additional threads, that is, monitor the CPU usage and application thread availability as well as the memory of the J2EE Engine after you have applied the change.

62

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

o

Important: Keep in mind that not all adapters are able to handle requests in parallel, even if you assign more threads on the Messaging System queues. In such a case adding additional threads will not speed up processing significantly. See chapter Adapter Parallelism for details.

o

A backlog of messages in the messaging system is highly critical for synchronous scenarios due to timeouts that might occur. Therefore, the number of synchronous threads always has to be tuned so that no bottleneck occurs. Often a backlog in synchronous queues is caused due to bad performance on the receiving backend system.

o

More information is provided in the blog Tuning the PI Messaging System Queues.

6.2.1

Messaging System in between AFW Sender Adapter and Integration Server (Outbound)

Open the Audit Log of a message and look for the following entry: “The message was successfully retrieved from the send queue”. Note that the queue name would be “Call Queue” if the message was synchronous. To determine how long the message waited to be picked up from the queue, compare the timestamp of the above entry with the timestamp for the step “Message successfully put into the queue”. A large time difference between those two timestamps would therefore indicate a bottleneck for consumer threads on the sender queues in the Messaging System.

Instead of checking the latency of single messages, we recommend using Wily Introscope to monitor the queue behavior as shown above. Asynchronous messages only use one thread for processing in the messaging system. Multiple threads are used for synchronous messages. The adapter thread puts the message into the messaging system queue and will wait till the messaging system delivers the response. The adapter thread is therefore not available for other tasks until the response is returned. A consumer thread on the call queue sends the message to the Integration Engine. The response will be received by a third thread (consumer thread of send queue) which correlates the response with the original request. After this, the initiating adapter thread will be notified to send the response to the original sender system. This correlation can be seen in the audit log of the synchronous message below.

63

SAP NETWEAVER 7.1 AND HIGHER

6.2.2

PERFORMANCE CHECK GUIDE

Messaging System in Between Integration Server and AFW Receiver Adapter (Inbound)

Open the Audit Log of a message and look for the following entry: “Message successfully put into the queue”. Compare this timestamp with the timestamp for the step “The message was successfully retrieved from the receive queue”. The time difference between these two timestamps is the time that the message waited in the messaging system for a free consumer thread. A large time difference between those two timestamps indicates a bottleneck in the adapter specific inbound queues of the Messaging System.

6.2.3

Interface Prioritization in the Messaging System

SAP NetWeaver PI 7.1 and higher offers the possibility for prioritization of interfaces similar to the ABAP stack. This feature is especially helpful if high volume interfaces run at the same time with business critical interfaces. To minimize the delay for critical messages, the Messaging System now makes it possible for you to define high, medium, and low priority processing at interface level. Based on the priority, the Dispatcher Queue of the messaging system (which will be the first entry point for all messages) will forward the messages to the standard adapter-specific queues. The priority assigned to an interface determines the number of messages that are forwarded once the adapter-specific queues have free consumer threads available. This is done based on a weighting of the messages to be reloaded. The weights for the different priorities are as follows: High: 75, Medium: 20, Low: 5. Based on this approach you can ensure that more resources can be used for high priority interfaces. The screenshot below shows the UI for message prioritization available in pimon  Configuration and Administration  Message Prioritization:

64

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

The number of messages per priority can be seen in a dashboard in Wily as shown below:

You can find more details on configuring usage prioritization within the AAE at: SAP Online Help navigate to SAP NetWeaver  SAP NetWeaver PI  SAP NetWeaver Process Integration 7.1 Including Enhancement Package 1  SAP NetWeaver Process Integration Library  Function-Oriented View  Process Integration  Process Integration Monitoring  Component Monitoring  Prioritizing the Processing of Messages.

6.2.4

Avoid Blocking Caused by Single Slow/Hanging Receiver Interface

As already discussed above, some receiver adapters process messages sequentially on one server node by default. This is independent of the number of consumer threads defined for the corresponding receiver queue in the messaging system. For example, even though you have configured 20 threads for the JDBC Receiver, one Communication Channel will only be able to send one request to the remote database at a given time. If there are many messages for the same interface, all of them will get a thread from the messaging system but will be blocked when performing the adapter call. In the case of a slow message transmission, for example, for EDI interfaces using ISDN technology or remote system with small network bandwidth, this behavior can cause a “hanging” situation for a specific adapter since one interface can block all messaging threads. As a result, all the other interfaces will not get any resources and will be blocked. 65

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Based on this, a parameter messaging.system.queueParallelism.maxReceivers was introduced which can be set in NWA in service XPI Service: Messaging System. With this parameter you can restrict the maximum number of receiver consumer threads that can be used by one interface. In older releases (lower 7.31 SP11 and 7.4 SP6) this is a global parameter at the moment that affects all adapters. It should therefore not be set too restrictively. If you need high parallel throughput one option is to set the maxReceiver parameter to 5 (so that each interface can use 5 consumer threads for each server node) and increase the overall number of threads on the receive queue for the adapter in question (for example, JDBC) to 20. With this configuration it will be possible for four interfaces to get resources in parallel before all threads are blocked. For more information, see Note 1136790 - Blocking Receiver Channel May Affect the Whole Adapter Type and SDN blog Tuning the PI Messaging System Queues. In the screenshot below you see the impact of maxReceivers parameter when a backlog for one interface occurs. Even though there are more free SOAP threads available they are not consumed. Hence the free SOAP threads could be assigned to other interfaces running in parallel.

This parameter has a direct impact on the queue where message backlogs occur. As mentioned earlier usually a backlog should occur on the Dispatcher queue only since it is responsible to dispatch threads only if there are free consumer threads in the adapter specific queue. When setting maxReceiver you restrict the threads per interface and this means that less consumer threads are blocked. When one interface is facing a high backlog there will always be free consumer threads and therefore the dispatcher queue can dispatch the message to the adapter specific queue. Hence the backlog will appear in the adapter specific queue so that the message prioritization would not work properly any longer:

66

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Per default the maxReceiver parameter is only relevant for asynchronous message processing (ICo and classical dual-stack scenarios). The reason here is that a wait time for synchronous scenarios should be avoided by all means and therefore additional restrictions in the number of available threads can be very critical. Therefore it is usually advisable to not limit the threads per interface but increase the overall number of available threads. In case you have many high volume synchronous scenarios with different priority that run in parallel it might be advisable to limit the threads each interface can use. To do so you have to set the parameter messaging.system.queueParallelism.queueTypes to Recv, ICoAll as described in SAP Note 1493502 - Max Receiver Parameter for Integrated Configurations. Enhancement with 7.31 SP11 and 7.4 SP6: With Note 1916598 - *NF* Receiver Parallelism per Interface an enhancement was introduced that allows the specification of the maximum parallelization on more granular level. This new feature has to be activated by setting the parameter messaging.system.queueParallelism.perInterface to true. Using a configuration UI you can specify rules to determine the parallelization for all interfaces of a given receiver service. If no rule for a given interface is specified the global maxReceivers value will be considered. If the receiver service corresponds to a technical business system this configuration would help to restrict the parallel requests from PI to that system. This also works across protocols so that e.g. it could be used to restrict the number of parallel requests using Proxy or IDoc to the same ERP receiver system. Below you can find a screenshot of the configuration UI in NWA  SOA  Monitoring:

With the improvement mentioned above, also the dispatching mechanism in the dispatcher queue is changed so that it is aware of the maxReceiver settings. This means that now again the backlog will now be placed in the dispatcher queue and the prioritization will work properly.

6.2.5

Overhead based on interface pattern being used

Also the interface pattern configured can cause some overhead. When choosing the interface pattern “Stateless Operation” each message is parsed against its data type during inbound processing. This can cause performance and memory overhead but gives a syntactical check of the data received. When choosing “Stateless (XI30-Compatible)” no check of the data type is performed and therefore the overhead is avoided. Therefore for interface with good data quality and high message throughput this interface pattern should be chosen.

67

SAP NETWEAVER 7.1 AND HIGHER

6.3

PERFORMANCE CHECK GUIDE

Performance of Module Processing

If your analysis in chapter Adapter Performance Problem pointed to a problem in the Module Processing, then you must analyze the runtime of the modules used in the Communication Channel. Adapter modules can be combined so that one CC calls multiple modules in a defined sequence. Adapter modules can be custom-developed or SAP standard and can be used for many different purposes – like to transform the EDI message before sending it to the partner or to change header attributes of a PI message before forwarding it to a legacy application. Due to the flexible usage of adapter modules there is also no standard way to approach performance problems. In the audit log shown below you can see two adapter modules. One module is a customer-developed module called SimpleWaitModule. The next module called CallSAPAdapter, which is a standard module that inserts the data into the messaging system queues.

Module Processing

In the audit log you will get a first impression of the duration of the module. In the example above you can see that the SimpleWaitModule requires 4 seconds of processing time. Once again, Wily Introscope offers a better overview of the processing time of modules. In the screenshot below you can see a dashboard showing the cumulative/average response times and number of invocations of different modules. If there is one that has been running for a very long time, then it would be very easy to identify since there will be a line indicating a much higher average response time. The tooltip displays the name of the module.

68

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Once you have identified the module with the long running time you have to talk to the responsible developer to understand why it is taking that long. Eventually the module executes a look-up to a remote system using JCo or JDBC which could be responsible for the delay. In the best case the module will print out information in addition to the audit log to detect such steps. If not, use the Wily Introscope transaction trace as explained in appendix Wily Transaction Trace. 6.4

Java only scenarios / Integrated Configuration objects

From SAP NetWeaver PI 7.1 on you can configure scenarios that only run on the Java stack using the Advance Adapter Engine (AAE) when only Java-based adapters are used. A new configuration object is used for this that is called Integrated Configuration. When using this, the steps executed so far in the ABAP pipeline (Receiver Determination, Interface Determination and Mapping) are also executed by the services in the Adapter Engine. 6.4.1

General performance gain when using Java only scenarios

The major advantage of AAE processing is a reduced overhead due to the context switches between ABAP and Java. Thus, the overall throughput can be increased significantly and the overall latency of a PI message can be reduced greatly. If possible, interfaces should always use the Integrated Configuration to achieve best performance. Therefore, the best tuning option is to change a scenario that is using Java based sender and receiver adapters from a classical ABAP-based scenario to an AAE-based one.

69

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Based on SAP internal measurements performed on 7.1 releases the throughput as well as the response time could be improved significantly as shown in the diagrams below.

70

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Based on these measurements the following statements can be made: 1) Significant performance improvements can be achieved always with local processing (if available) for a certain scenario. 2) This is valid for all adapter types; huge mapping runtimes, slow networks, slow (receiver) applications reduce the throughput and therefore, rated for the overall scenario, also reduce the benefit of local processing in terms of overall throughput and response time measurements. 3) Greatest benefit is recognized for small payloads (comparison: 10k, 50k, 500k) and asynchronous messages. 6.4.2

Message Flow of Java only scenarios

All the Java based tuning options mentioned in chapter Analyzing the (Advanced) Adapter Engine and Long Processing Times for “PLSRV_MAPPING_REQUEST” are still valid. Just the calling sequence is different. The following describes the steps used in Integrated Configuration to help you better understand the message flow of an interface. The example is JMS to Mail: 1) Enter the JMS sender adapter 2) Put into the dispatcher queue of the messaging system 3) Forwarded to the JMS send queue of the messaging system 4) Message is taken by JMS send consumer thread: a. No message split used: In this case, the JMS consumer thread will do all the necessary steps previously done in the ABAP pipeline (like receiver determination, interface determination, and mapping) and will then also transfer the message to the backend system (remote database, in our example). Thus, all the steps are executed by one thread only. b. Message split used (1:n message relation): In this case there is a context switch. The thread taking the message out of the queue will process the message up to the message split step. It will create the new message and put it in the Send queue again from where it will be taken by different threads (which will then map the child message and finally send it to the receiving system) As we can see in this example for Integrated Configuration, only one thread does all the different steps of a message. The consumer thread will not be available for other messages during the execution of these steps. The tuning of the Send queue (Call for synchronous messages) is therefore much more important for scenarios using Integrated Configuration than for ABAP-based scenarios. The different steps of the message processing can be seen in the audit log of a message. If, for example, a long-running mapping or adapter module is indicated, then you can use the relevant chapter of this guide. There is no difference in the analysis except that for mappings no JCo connection is required since the mapping call is done directly from the Java stack.

71

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

The example audit logs below are based on a Java only synchronous SOAP to SOAP scenario.

In the highlighted areas you can see that all the steps are very fast except the mapping call which lasts around 7 seconds. To analyze the long duration of the mapping now the same steps as discussed in chapter Long Processing Times for “PLSRV_MAPPING_REQUEST” have to be followed.

72

SAP NETWEAVER 7.1 AND HIGHER

6.4.3

PERFORMANCE CHECK GUIDE

Avoid blocking of Java only scenarios

Like classical ABAP based scenarios, Java only scenarios can face backlog situations in case of a slow or hanging receiver backend. Since the Java only interfaces are only using send queues the restriction of consumer threads on the receiver queue as described in chapter Avoid Blocking Caused by Single Slow/Hanging Receiver Interface is no solution. Based on that an additional property messaging.system.queueParallelism.queueTypes of service SAP XI AF MESSAGING was introduced via Note 1493502 - Max Receiver Parameter for Integrated Configurations. The parameter can be set for synchronous and asynchronous interfaces. Please keep in mind that messaging.system.queueParallelism.maxReceivers is a global parameter for all adapters and for all configured Quality of Service. This global restriction can be avoided starting with 7.31 SP11/ 7.40 SP06, as per SAP Note 1916598 - *NF* Receiver Parallelism per Interface. Usually a restriction for the parallelization can be highly critical for synchronous interfaces. Therefore we generally recommend to set messaging.system.queueParallelism.queueTypes in most cases to Recv, IcoAsync only. 6.4.4

Logging / Staging on the AAE (PI 7.3 and higher)

Up to PI 7.11 on the PI Adapter Engine only one message version was persisted. Especially for Java only scenarios this was often not sufficient for troubleshooting since for instance the result of a mapping could not be verified. Starting from PI 7.3 the persistence steps can be configured globally in the Adapter Engine for synchronous and asynchronous interfaces. This is similar to the LOGGING or LOGGING_SYNC on the ABAP stack. In later version of PI you will be able to do the configuration on interface level. The versions that can be persisted are shown in the diagram below (the abbreviations in green are the values for the configuration):

73

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

In the Messaging System we generally distinguish Staging (versioning) and Logging. An overview is given below:

For details about the configuration please refer to the SAP online help Saving Message Versions. Similar to the LOGGING on the ABAP Integration Server, persisting several versions of the Java message can cause a high overhead on the DB and can cause a decrease in performance. The persistence steps can be seen directly in the audit log of a message:

Also the Message Monitor shows the persisted versions:

74

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

While this additional monitoring attributes are extremely helpful for troubleshooting they should be used carefully from a performance perspective. Especially the number of message versions and the logging/staging mode you choose can have a severe impact on performance. Therefore you have to find the balance between business requirement and performance overhead. Some guidelines on how to use staging and logging are summarized in SAP Note 1760915 - FAQ: Staging and Logging in PI 7.3 and higher. 6.5

J2EE HTTP load balancing

With NetWeaver version 7.1 and higher the Java HTTP load balancing is no longer done by the Java Dispatcher but by the ICM. The default load balancing rules of the ICM are designed with a stateful application in mind giving priority to the balancing of HTTP sessions. These default rules are not optimal for stateless applications such as SAP PI. In the past we could observe an unequal load distribution across Java server nodes caused by this. In case of high backlogs this can cause a delay in the overall message processing of the interface because one server node has more messages assigned than others and therefore not all available resources are used. A typical example can be seen in the Wily screenshot below:

SAP Note 1596109 – Uneven distribution of HTTP requests on Java server nodes introduces new load balancing rules for stateless applications like PI. Please follow the description in the Note to ensure the messages are distributed equally across the available server nodes. In the meantime these load balancing rules can also be executed using the CTC templates as described in Notes 1756963 and 1757922. Also, this has been included in the PI Initial Setup wizard in order to execute this task automatically as a postinstallation step (see Note 1760700 - PI CTC: Add HTTP loadbalancing to initial setup).

75

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

For the example given above we could see a much better load balancing after the new load balancing rules were implemented. This can be seen in the following screenshot:

Please Note: The load balancing rules mentioned above are only responsible to balance the messages across the available server nodes of one instance. HTTP load balancing across application servers is done by the SAP Web Dispatcher. More information about this is provided in the Guide How To Scale SAP PI 7.1. 6.6

J2EE Engine Bottleneck

All tuning that can be done in the AFW is limited by the ability of the J2EE Engine to handle a given amount of threads and to provide enough memory that is needed to process all requests. Of course, the CPU is also a limiting factor, but this will be discussed in Chapter 9.1. From SAP NetWeaver PI 7.1 onwards, the SAP VM is used on all platforms. Therefore the analysis of all platforms can use the same tools. 6.6.1

Java Memory

The first analysis of the J2EE memory is usually done using the garbage collection (GC) behavior of the Java Virtual Machine (JVM). To get the information of the GC written to the standard log file, the JVM has to be started with the parameter –verbose:gc.  Analyze the file std_server.out for the frequency of GCs, their duration, and the allocated memory.  Look for unusual patterns indicating an Out-of-Memory error or an unintentional restart of your J2EE engine. This would be visible if the allocated memory drops to 0 at a given point in time. A healthy Garbage Collection output would display a saw tooth pattern – that is, the memory usage would increase over time but then go down to its original value as soon as a full Garbage Collection is triggered. You should see the memory usage go down to initial values during low volume times (nighttime).  Pay attention to GCs with a high duration. This is important because no PI message can be mapped or processed by the J2EE Adapter Framework during a GC in a JAVA VM. One of the many reasons for long GCs that should be mentioned here is paging (see chapter 9.2). If the heap space of the J2EE Engine is exhausted, all objects in the heap are evaluated for their references. If there is still a reference, the object is kept. If there is no reference, the object is deleted. If some or all of the objects have been swapped out due to insufficient RAM, then they have to be swapped in to be evaluated. This leads to very long runtimes for GCs. GCs of more than 15 minutes have been observed on swapping systems.

76

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

 Different tools exist for the Garbage Collection Analysis: 1) Solution Manager Diagnostics: Solution Manager Diagnostics (SMD) offers a memory analysis application that is able to read the GC output in the std_server.out files. To start, call the Root Cause Analysis section. In the Common Task list you find the entry Java Memory Analysis which will give you the chance to upload your std_server.out file. It shows the memory allocated after the GC and also prints the duration of each GC (black dots with duration scale on right axis). A normal GC output should show a saw tooth pattern where the memory always goes down to an initial value (especially during low volume times). This is shown in the example screenshot below.

2)

Wily Introscope: Again, Wily Introscope offers different dashboards that can be used to check the Garbage Collection behavior on the J2EE engine. The dashboard can be found using the J2EE Overview  J2EE GC Overview and shows important KPIs for the GC, such as the count of GC or the duration for each interval. The %GC Time dashboards shows the ratio of time spent in GC. An example is shown in the screenshot below.

77

SAP NETWEAVER 7.1 AND HIGHER

3)

PERFORMANCE CHECK GUIDE

Netweaver Administrator: The NWA also offers a view to monitor the Java heap usage. However, the important information about the duration of the GC is not available. This monitor can be found by navigating to Availability and Performance Management  Java System Reports  Choose Report System Health.

In NetWeaver 7.3 you find this data in Availability and Performance Management  Resource Monitoring  History Reports. There you can build your own Memory report of the monitoring data provided. The screenshot below e.g. shows the Used Memory per server node.

78

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

4) SAP JVM Profiler: The SAP JVM profiler allows the analysis of the extensive GC information stored in the sapjvm_gc.prf files of the server node folders or to directly connect to the server node to retrieve this information. More information can be found at http://wiki.scn.sap.com/wiki/display/ASJAVA/Java+Profiling:

79

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Procedure: o

Search for restarts of the J2EE Engine by searching for the string “is starting” in the file std_server.out. Apart from the first of these entries (which simply marks the initial start of the J2EE Engine), the second and any later entries mark a fatal situation after which the J2EE server node had to be restarted. This could, for example, be an out–of-memory situation. Search the log above those entries for a possible reason.

o

Search for exceptions and fatal entries. Was an out-of-memory error reported (search for pattern OutOfMemory)? Was a thread shortage reported?

o

This document does not deal with tuning the J2EE Engine. If at this point it becomes clear that the J2EE Engine is the limiting factor, proceed with SAP Notes or open a customer incident. Only some very basic recommendations are given here.

o

The parameters for the SAP VM are defined by Zero Admin templates. From time to time new Zero Admin templates might be released which change important parameters for the SAP VM. Thus, you should check SAP Note 1248926 - AS Java VM Parameters for NetWeaver 7.1-based Products regularly and apply the changes accordingly.

o

The problem might be caused by too high parallelization when processing large messages. This should be restricted to avoid overloading of the Java heap memory. 1) In case the interfaces are using the ABAP Integration Server use the large message queue filters to restrict the parallelization of mapping calls from ABAP queues. To do so set parameter EO_MSG_SIZE_LIMIT to e.g. 5.000 to direct all large messages to dedicated XBTL* or XBTM* queues. 2) To restrict the parallelization on the Java Messaging System you can use the large message queue mechanism described in chapter Large message queues on the Messaging System.

o

If it is already obvious that the memory of your J2EE Engine is getting short and nothing can be changed on the scenarios itself you have several options for scaling your PI landscape. Especially for large PI installations with high message throughput or large messages SAP Note 1502676 - Scaling up large PI installations describes potential options. 1) Increase the Java Heap of your server node: If sufficient physical memory is available, increase the default heap size of 2 GB to a higher value such as 3 or 4 GB. Experience shows that the SAP VM can handle these heap sizes without major impact on performance. Larger heap sizes can cause longer processing time of GCs, which in turn can affect the PI application negatively. Increasing the maximum heap size on the J2EE engine will automatically adapt the new area of the heap due to a dynamic configuration (in newer J2EE versions this will be per default set to 1/6 of the overall heap). After increasing the heap size, monitor the duration of the GC execution. 2) Add an additional server node to distribute the load over more processes. SAP recommends that productive systems have at least 2 Java server nodes per instance. Prerequisite is that your physical host has enough RAM and CPU. To minimize the overhead for J2EE cluster communication it is in general recommended configuring larger server nodes (as described above) then a high number of server nodes. 3) Add an additional application server consisting of a double stack Web Application Server (ABAP and J2EE Engine) to distribute the load to an additional hardware. Find details on the scaling of PI with multiple instances in the Guide How To Scale up SAP NetWeaver Process Integration. 4) Use a non-central Adapter Engine to separate large messages that cause the memory problem from business critical scenarios.

6.6.2

Java System and Application Threads

A system or application thread is required for every task that the J2EE Engine executes. Each of these thread types are maintained using a thread pool. If the pool of available threads is expired, no more requests can be processed. It is therefore important that you have enough threads available at all times.

80

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

o

In general, SAP recommends that you increase the system and application thread count for a PI system. As described in SAP Note 937159 - XI Adapter Engine is stuck, the application threads should be increased to 350 and the system threads to ‘120’ for all J2EE server nodes.

o

There are two options for checking the thread usage:

1) Wily Introscope: Wily Introscope provides a dashboard in the J2EE Overview section called J2EE CPU and Memory Detail that can also be used to monitor the thread usage on a PI engine. Different views exist for Application and System Threads as shown in the screenshot below.

2) NetWeaver Administrator (NWA): The NWA also offers a monitor similar to the one available for the memory mentioned above. This monitor can be found by navigating to Availability and Performance Management  Java System Reports  Choose Report System Health and looking at the windows shown in the screenshot below.

Again in NetWeaver 7.3 and higher the monitor can be found at Availability and Performance Management  Resource Monitoring  History Report. An example for the Application Thread Usage is shown below:

81

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

 Check if the ThreadPoolUsageRate is below 80% for both the application threads and the system threads. Also check the maximum value by choosing a longer history in Wily Introscope  Has the thread usage increased over the last days/weeks?  Is there one server node that shows a higher thread usage than others? 3) SAP Management Console: The SAP Management Console is an essential tool to analyze different aspects of the J2EE engine. It also offers a thread dump view, similar to SM50 in ABAP, showing the activity of all the threads of a given Java Application Server. More information about the SAP MC can be found in Note 1014480 SAP Management Console (SAP-MC) and the online help. The SAP Management Console can be started as Java web-start application using http://:513. 

Navigate to AS Java  Threads and check for any threads in red status (>20 secs running time). This indicates long running threads and you should check for the related thread stacks. The example below shows a long running thread related to the PI directory. You can see the user doing the request and can see that the thread is waiting for an HTTP response from the CPACache:

82

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

 If you are not sure what is causing the problem you should take multiple thread dumps (at least 3 every 30 seconds) for all server nodes. This can be done in the Management Console navigating to AS Java  Process Table.

The Thread Dump Viewer tool (TDV) from Note 1020246 - Thread Dump Viewer for SAP Java Engine can be used to analyze the produced thread dumps (inside the results archive that can be downloaded to your local PC). 6.6.3

FCA Server Threads

An additional type of thread was introduced with SAP NetWeaver PI 7.1 that is responsible for HTTP traffic: FCA Server Threads. The FCA Server Threads are responsible for receiving HTTP calls at the Java side (after the Java dispatcher is no longer available). Also, FCA Threads use a Thread Pool. Fifteen FCA Threads are configured by default but based on SAP Note 1375656 - SAP Netweaver PI System Parameters we recommend that you increase it to 50. This can be done in the NWA by changing the parameter FCAServerThreadCount of service HTTP Provider. FCA Server Threads are particularly crucial for synchronous message transfer and for HTTP-based scenarios like WebServices or calls from the ABAP engine to the Adapter Engine. If there is a long duration of the HTTP call due to a slow backend system then the thread is blocked for the whole time and not available to send other HTTP requests (other PI messages).

83

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

The configured FCAServerThreadCount represents the maximum number of threads working on a single entry point. An entry point in the PI sense could e.g. be the SOAP Adapter Servlet (share with all channels as described in Tuning SOAP sender adapter). o

If response times for specific entry point are high (>15 seconds) additional FCA threads will be spawned to be available for parallel HTTP based incoming requests using different entry points only. This ensures that in case of problems with one application not all other applications will be blocked constantly.

o

An overall maximum of 20xFCAServerThreadCount may be used in the J2EE engine

There is currently no standard monitor available for FCA Threads except the thread view in the SAP Management Console. A new dashboard will be integrated into Wily Introscope soon and will also show the FCA Server Threads that are in use.

6.6.4

Switch Off VMC

The Virtual Machine Container (VMC) is enabled by default for an ABAP WebAS 7.1. Since PI does not use VMC at runtime, the VMC can be switched off on a PI system to avoid resource overhead. This recommendation is based on SAP Note 1375656 - SAP NetWeaver PI System Parameters. You can activate the VM Container setting the profile parameters vmcj/enable = off which can be changed using profile maintenance (transaction RZ10). Once you have made the change, you need to restart the SAP Web Application Server instance.

84

SAP NETWEAVER 7.1 AND HIGHER

7

PERFORMANCE CHECK GUIDE

ABAP PROXY SYSTEM TUNING

Every ABAP WebAS > 6.20 includes an ABAP Proxy runtime. This enables a SAP system to talk native PI protocol (XI-SOAP) so that no costly transformation is necessary. In general, ABAP proxy is just a framework that is triggered by (sender) or triggers (receiver) applicationspecific coding. It is not possible to give a general tuning recommendation because the applications and use cases of ABAP proxy can differ so greatly. In this section we would like to highlight the system tuning options that can be applied to improve the throughput at the ABAP Proxy backend side. In general you can increase the throughput of EO interfaces by changing the queue parallelization. Two different parameters exist to change the number of qRFC queues on the ABAP Proxy backend. As in PI, ABAP Proxy processing is based on qRFC inbound queues (SMQ2) with specific names. A sender proxy uses XBTS* queues (10 parallel by default) and a receiver proxy XBTR* (20 parallel by default). The sender proxy only forwards the message to PI by executing an HTTP call that must not take too much time and which is not resource-critical. On the other hand, the receiver proxy executes the inbound processing of the messages based on the application context (and which can be very time-consuming). It is therefore more likely to face a backlog situation in the Receiver Proxy queues. In the screenshot below you can see the performance header of a receiver proxy message that required around 20 minutes in the PLSRV_CALL_INBOUND_PROXY (which corresponds to the posting of the application data).

Of course, such a long-running message will block the queue and all messages behind it will face a higher latency. Since this step is purely application-related, it is only possible to perform tuning at the application side. The number of sender and receiver ABAP Proxy queues can be changed in transaction SXMB_ADM on the proxy backend with parameter EO_INBOUND_PARALLEL and sub parameter SENDER (XBTS* queues) and RECEIVER (XBTR* queues). The prerequisite for increasing the number of queues is that there are enough qRFC resources are available at the proxy system (as discussed in section qRFC Resources (SARFC)). Also, as described in chapter Message Prioritization on the ABAP Stack, prioritization can be configured in the ABAP Proxy system. This can be used to separate runtime-critical interfaces from interfaces with a long application processing (as shown above).

85

SAP NETWEAVER 7.1 AND HIGHER

7.1

PERFORMANCE CHECK GUIDE

New enhancements in Proxy queuing

The queuing mechanism was enhanced for sender and receiver proxy systems with Note 1802294 (Receiver) and Note 1831889 (Sender). After implementing these two Notes interface specific queues will be used and tuning of these queues on interface level will be possible. This can be very helpful in cases where one receiver interface shows very long posting time in the application coding that cannot be further improved. Other messages for more business critical interfaces will eventually be blocked by this message due to the common usage of the XBTR* queues. On the sender side of a Proxy system the processing of the queues call the central PI hub. In general the processing time there should be fast. But in case of high volume interfaces you might want to slow down less business critical interfaces to avoid overloading of your central PI hub. The prioritization mechanisms available prior to these Notes did not provide such options. This is achieved by adding an interface specific identifier to the queue-name. In the screenshot below you see a comparison in the queue names for the old framework (red) and the new framework (blue):

For the sender queues the following new parameters are introduced o

Parameter EO_QUEUE_PREFIX_INTERFACE with sub parameter SENDER/SENDER_BACK: Should be set to ‘1’ to active interface specific queue. This is a global switch affecting all sender interfaces.

o

Parameter EO_INBOUND_PARALLEL_SENDER without sub parameter determines the general number of queues per interface  Using a sub parameter the parallelization of individual interfaces can be changed. This can e.g. be used to assign more queues for high priority interfaces  Wildcards can be used in the to group interfaces sharing e.g. the same service name or the same interface name but different services.

For the receiver queues the following new parameters are introduced o

Parameter EO_QUEUE_PREFIX_INTERFACE with sub parameter RECEIVER: Should be set to ‘1’ to active interface specific queue. This is a global switch affecting all receiver interfaces.

o

Parameter EO_INBOUND_PARALLEL_RECEIVER without sub parameter determines the general number of queues per interface  Using a sub parameter the parallelization of individual interfaces can be changed. This can e.g. be used to assign more queues for high priority interfaces  Wildcards can be used in the to group interfaces sharing e.g. the same service name or the same interface name but different services.

This new feature also replaces the currently existing prioritization since it is in general more flexible and powerful. We therefore recommend adjusting existing prioritization rules using the new queuing possibilities described above.

86

SAP NETWEAVER 7.1 AND HIGHER

8

PERFORMANCE CHECK GUIDE

MESSAGE SIZE AS SOURCE OF PERFORMANCE PROBLEMS

The message size directly influences the performance of an interface. The size of a PI message depends on two elements: The PI header elements with a rather static size and the payload which can vary greatly between interfaces or over time for one interface (for example, larger messages during year-end closing). The size of the PI message header can cause a major overhead for small messages of only a few kB and can cause a decrease in the overall throughput of the interface. Furthermore, many system operations (like context switches or database operations) are necessary for only a small payload. The larger the message payload, the smaller the overhead due to the PI message header. On the other hand, large messages require a lot of memory on the Java stack which can cause heavy memory usage on ABAP or excessive garbage collection activity (see section Java Memory) that will also reduce the overall system performance. Very large messages can even crash the PI system by causing an Out-of-Memory exception, for example. You therefore have to find a compromise for the PI message size. Below you see throughput measurements performed by SAP. In general, the best throughput was identified for messages sizes of 1 to 5 MB.

The message size in these measurements corresponds to the XML message size processed in PI and not the size of the file or IDoc being sent to PI. You can use the Runtime Header of the ABAP stack to check the message size. Below you can see an example of a very small message. While the MessageSizePayload field describes the size of the payload in bytes (here 433 bytes), the MessageSizeTotal describes the total message size (header + payload). In the example this is around 14 KB demonstrating the overhead by the PI header for small messages. The next two lines describe the payload size before and after the mapping. In the example below, the mapping reduces the payload size. The last two lines determine the size of the response message that is sent back to PI before and after the response mapping for synchronous messages.

87

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Based on the above observations, we highly recommend that you use a reasonable message size for your interfaces. During the design and implementation of the interface we therefore recommend using a message size of 1 to 5 MB, if possible. Messages can be collected at the sender side or by using IDoc packaging as described in Tuning the IDoc Adapter to achieve this. In case of large messages, a split has to be performed by changing the sender processing or by using the split functions available in the structure conversion of the File adapter. 8.1

Large message queues on PI ABAP

In case the interfaces are using the ABAP Integration Server use the large message queue filters to restrict the parallelization of mapping calls from ABAP queues. To do so set the parameter EO_MSG_SIZE_LIMIT of category TUNING to e.g. 5.000 to direct all messages larger 5 MB to dedicated XBTL* or XBTM* queues. The value of the parameter depends on the number of large messages and the acceptable delay that might be caused due to a backlog in the large message queue. To reduce the backlog the number of large message queues can also be configured via parameter EO_MSG_SIZE_LIMIT_PARALLEL of category TUNING. The default value is 1 so that all messages larger than the defined threshold will be processed in one single queue. Naturally the parallelization should not be set higher than 2 or 3 to avoid overloading of the Java memory due to parallel large message requests. 8.2

Large message queues on PI Adapter Engine

SAP Note 1727870 - Handling of large messages in the Messaging System introduces large message queues (virtual queues, no adapter behind) also for the Java based Adapter Engine. Contrary to the Integration Engine, it is not the size of a single large message only that determines the parallelization. Instead the sum of the size of the large messages across all adapters on a given Java server node is limited to avoid overloading the Java heap. This is based on so called permits that define a threshold of a message size. Each message larger than the permit threshold is considered as large message. The number of permits can be configured as well to determine the degree of parallelization. Per default the permit size is 10 MB and 10 permits are available. This means that large messages will be processed in parallel if 100 MB are not exceeded. To show this let us look at an example using the default values. Let us assume we have 5 messages waiting to be processed (status “To Be Delivered” on one server node. Message A has 5 MB, message B has 10 MB, message C has 50 MB, message D 150 MB, message E 50 MB and message F 40 MB. Message A is not considered large since the size is smaller than the permit size and is not considered large and can be immediately processed. Message B requires 1 permit, message C requires 5. Since enough permits are available processing will start (status DLNG). Hence for message D all available 10 permits would be required. Since the permits are currently not available it cannot be scheduled. If blacklisting is enabled the message will be put to error status (NDLV) since it exceed the maximum number of defined permits. In that case the message would have to be restarted manually. Message E requires 5 permits and can also not be scheduled. But since there are 4 permits left message F is put to DLNG. Due to the smaller size message B and message F finish first releasing 5 permits. This is sufficient to schedule message E which requires 5

88

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

permits. Only after message E and C have finished message D can be scheduled consuming all available permits. The example above shows a potential delay a large message could face due to the waiting time for the permits. But the assumption is that large messages are not time critical and therefore additional delay is less critical than potential overload of the system. The large message queue handling is based on the Messaging System queues. This means that restricting the parallelization is only possible after the initial persistence of the message in the Messaging System queues. Per default this is only done after the Receiver Determination. Therefore if you have a very high parallel load of incoming large requests this feature will not help. Instead you would have to restrict the size of incoming requests on the sender channel (e.g. file size limit in the file adapter or the icm/HTTP/max_request_size_KB limit in the ICM for incoming HTTP requests). If you have very complex extended receiver determination or complex content based routing it might be useful to configure staging in the first processing step of the Messaging System (BI=3) as described in Logging / Staging on the AAE (PI 7.3 and higher). The number of permits consumed can be monitored in PIMON  Monitoring  Adapter Engine Status. The number of threads corresponds to the number of consumed permits:

In newer Wily versions there is also a dashboard showing the number of consumed permits as well. The Worker Threads on the right hand side of the screenshot correspond to the permits:

89

SAP NETWEAVER 7.1 AND HIGHER

9

PERFORMANCE CHECK GUIDE

GENERAL HARDWARE BOTTLENECK

During all tuning actions discussed in the previous chapters you must keep in mind that the limitation of all activities is set by the underlying CPU and memory capacity. The physical server and its hardware have to provide resources for three PI runtimes: the Adapter Engine, the Integration Engine, and the Business Process Engine. Tuning one of the engines for high throughput leaves fewer resources for the remaining engines. Thus, the hardware capacity has to be monitored closely. 9.1

Monitoring CPU Capacity

When monitoring the CPU capacity it is not enough to use the average that is displayed in the “Previous Hours” section of transaction ST06. Instead, you should start your interface and monitor the CPU while it is running. The best procedure is to use operating system tools to monitor the CPU usage (particularly if using hardware virtualization). The SMD Host Agent also reports CPU data to Wily Introscope which can be viewed using dashboards, as shown below. This example shows two systems where one is facing a temporary CPU overload and the other a permanent one.

Also the NWA offers a view on the CPU activity from NetWeaver 7.3 on via Availability and Performance Management  Resource Monitoring  History Reports. There you can build your own report based on the “CPU utilization” data as shown in the screenshot:

90

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

You can also monitor the CPU usage in ST06 as described below: Procedure Log on to your Integration Server and call transaction ST06. Take a look at the snapshot that is provided on the entry screen, than navigate to “Detailed Analysis Menu”.  Snapshot view: Is the idle time (upper left corner) at a reasonable value at all times, for example, around 10% or higher?  Snapshot view: The load average is a good indicator of the dimension of the bottleneck. It describes the number of processes for each CPU that are in a wait queue before they are assigned to a free CPU. As long as the average remains at one process for each available CPU, the CPU resources are sufficient. Once there is an average of around three processes for each available CPU, there is a bottleneck at the CPU resources. In connection with a high CPU usage, a high value here can indicate that too many processes are active on the server. In connection with a low CPU usage, a high value here can indicate that the main memory is too small. The processes are then waiting due to excessive paging.  Detailed Analysis View  TOP CPU. Which thread uses up the highest amount of CPU? Is it the J2EE Engine, the work processes, or the database?

9.2

o

There are different follow-up actions to be taken depending on the findings of the second check: The first option is of course to simply increase the hardware. This should be accompanied by a thorough sizing (for which SAP provides the Quicksizer in the Service Marketplace http://service.sap.com/quicksizer ).

o

The second option is to identify the CPU-consuming threads and to think about a reduction of the load in this component. For example, if the J2EE Engine is the largest consumer then it is worth reevaluating the number of concurrent mapping threads and/or consumer queues as described in the previous chapters.

o

SAP Note 742395 - Analyzing High CPU Usage by the J2EE Engine provides a good entry point if the J2EE Engine consumes too much CPU.

o

If it is the work processes that consume a lot of CPU, use transaction ST03N to figure out if it is the Integration Server or the Business Process Engine. You can do that by looking at the CPU usage per user: BPE uses the user WF-BATCH while the Integration Server uses the last user who activated a qRFC queue (typically PIAFUSER or PIAPPLUSER). Use the previous chapter to restrict these activities. Monitoring Memory and Paging Activity

As stated above, paging is very critical for the Java stack since it influence the Java GC behavior directly. Therefore, paging should be avoided in every case for a Java-based system. The OS tools are the most reliable for monitoring the paging activity on your system. Transaction ST06 can also be used to perform a snapshot analysis. Take a look at the snapshot that is provided on the entry screen:  Snapshot view: Is there enough physical memory available?  Snapshot view: Is there considerable paging? Paging can have a negative influence on your J2EE Engine. If it does, this can be seen in long GC times as described in Chapter 6.6.1.

9.3

Monitoring the Database

Monitoring database performance is a rather complex task and requires information to be gathered from a variety of sources. Only a basic set of indicators is given for the purpose of this guide. A more detailed

91

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

analysis is required if these indicators point to a major performance problem in the database. If assistance is needed, open a support message under the component BC-DB-ORA (or MSS, SDB, DB6 respectively). 9.3.1

Generic J2EE database monitoring in NWA

The next sections of this chapter will be mainly based on ABAP transactions to do a detailed analysis of the database performance.The NWA, however, offers possibilities to monitor the database performance for the Java based tables. This is especially helpful for Java only AEX/PO or Non-Central AAEs. In the NWA you can use the “Open SQL Monitor” in the Troubleshooting section of NWA. The official documentation can be found in the Online Help. Monitoring DB via DBACOCKPIT transaction (ABAP based) is possible also for the Java-only systems from SAP Solution Manager (DBA Cockpit must be configured in the Managed System Setup). It will be too much to outline all the available functionalities here. Instead only a few key capabilities will be demonstrated. If you want to e.g. see the number of select, update or insert statements in your system you can use the “Table Statistics Monitor”. It clearly shows you which tables in your system are accessed the most frequently:

To see the processing time of the different statements on your system you can use the “Open SQL statistics” monitor. There you can see the count, the total, average and max. processing time of the individual SQL

92

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

statements and can therefore identify the expensive statements on your system:

Per default the recorded period is always from the last restart of the system. If you would like to just look at the statistics for a specific time period (e.g. during a test) you have the option of resetting the statistics in the individual monitor prior to the test.

9.3.2

Monitoring Database (Oracle)

Procedure Log on to your Integration Server and call transaction ST04.

93

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Many of the numbers in the screenshot above depend on each other. The following checklist names a few key performance figures of your Oracle database.  The data buffer quality, for instance, is based on the ratio of physical reads versus the total number of reads. The lower the ratio, the better the buffer quality. The data buffer quality should be better than 94%; the statistics should be based on 15 million total reads. This number of reads ensures that the database is in an equilibrated state.  Ratio of user and recursive calls. A good performance is indicated by ratio values greater than 2. Otherwise the number of recursive calls, compared to the number of user calls, is too high. Over time this ratio always declines because more and more SQL statements get parsed in the meantime.  Number of reads for each user call. If this value exceeds 30 blocks for each user call, this indicates an expensive SQL statement.  Check the value of Time/User call. Values larger than 15 ms often indicate an optimization issue.  Compare busy wait time versus CPU time. A ratio of 60:40 generally indicates a well-tuned system. Significantly higher values (for example, 80:20) indicate room for improvement.  The DD-cache quality should be better than 80%. 9.3.3

Monitoring Database (MS SQL)

Procedure To display the most important performance parameters of the database, call Transaction ST04, or choose Tools  Administration  Monitor  Performance  Database  Activity. An analysis is only meaningful if the database has been running for several hours with a typical workload. To ensure a significant database

94

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

workload, we recommend a minimum of 500 CPU busy seconds. Note: The default values displayed in section Server Engine are relative values. To display the absolute values, press button Absolute values. Check the values in (1).

1 > 6h 1 4 1 3 1

2 > 98 1

5 1

 The cache hit ratio (2), which is the main performance indicator for the data cache, shows the average percentage of requested data pages found in the cache. This is the average value since startup. The value should always be above 98 (even during heavy workload). If it is significantly below 98, the data cache could be too small. To check the history of these values, use Transaction ST04 and choose Detail Analysis Menu  Performance Database. A snapshot is collected every 2 hours.  Memory setting (5) shows the memory allocation strategy used, and shows the following:  FIXED: SQL Server has a constant amount of memory allocated, which is set by SQL Server configuration parameters min server memory (MB) = max server memory (MB).  RANGE: SQL Server dynamically allocates memory between min server memory (MB) < > max server memory (MB).  AUTO: SQL Server dynamically allocates memory between 4 MB and 2 PB, which is set by min server memory (MB) = 0 and max server memory (MB) = 2147483647.  FIXED-AWE: SQL Server has a constant amount of memory allocated, which is set by min server memory (MB) = max server memory (MB). In addition the address windowing extension functionality of Windows 2000 is enabled. 9.3.4

Monitoring Database (DB2)

Procedure To get an overview of the overall buffer pool usage, catalog, and package cache information, go to transaction ST04, choose Performance  Database  section Buffer Pool (or section Cache respectively).

95

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

 Buffer Pools Number: Number of buffer pools configured in this system.  Buffer Pools Total Size: The total size of all configured buffer pools in KB. If more than one buffer pool is used, choose Performance  Buffer Pools to get the size and buffer quality for every single buffer pool.  Overall buffer quality: This represents the ratio of physical reads to logical reads of all buffer pools.  Data hit ratio: In addition to overall buffer quality, you can use the data hit ratio to monitor the database: (data logical reads - data physical reads) / (data logical reads) * 100%  Index hit ratio: In addition to overall buffer quality, you can use the index hit ratio to monitor the database: (index logical reads - index physical reads) / (index logical reads) * 100%.  Data or Index logical reads: The total number of read requests for data or index pages that went through the buffer pool.  Data or Index physical reads: The total number of read requests that required I/O to place data or index pages in the buffer pool.  Data synchronous reads or writes: Read or write requests performed by db2agents.

96

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

 Catalog cache size: Maximum size of the catalog cache that is used to maintain the most frequently accessed sections of the catalog.  Catalog cache quality: Ratio of catalog entries (inserts) to reused catalog entries (lookups).  Catalog cache overflows: Number of times that an insert in the catalog cache failed because the catalog cache was full (increase catalog cache size).  Package cache size: Maximum size of the package cache that is used to maintain the most frequently accessed sections of the package.  Package cache quality: Ratio of package entries (inserts) to reused package entries (lookups).  Package cache overflows: Number of times that an insert in the package cache failed because the package cache was full (increase package cache size). 9.3.5

Monitoring Database (MaxDB / SAP DB)

Procedure As with the other database types, call transaction ST04 to display the most important performance parameters.

97

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

With the Database Performance Monitor (transaction ST04) and its submonitors, the CCMS in the SAP R/3 System allows you to view, in the SAP R/3 System, all of the information that can be used to identify bottleneck situations.  The SQL Statements section provides information about the number of SQL statements executed and related sizes.  The I/O Activity section lists physical and logical read and write accesses to the database.  The Lock Activity section (in transaction ST04) allows you to identify possible bottlenecks in the assignment of locks by the database.  The Logging Activity section combines information about the log area.  The Scan and sort activity section can be helpful in identifying that suitable indexes are missing.

 The Cache Activity section provides information about the usage of the caches and the associated hit rates. o

The data cache Hit rate should be 99% in a balanced system. If this is not the case, you need to check if the low hit rate is due to the size of the data cache being too small, or due to an unsuitable SQL statement. If necessary, you can increase the data cache by increasing the MaxDB parameter Data_Cache_Size (lower than 7.4) or Cache_Size (7.4 or higher).

o

The catalog cache hit rate should be 85% in a balanced system. It is not a cause for concern if the catalog hit rate is temporarily lower, since accesses to the system tables and an active command monitor can temporarily impair the hit rate. If the catalog cache is busy, the pages are paged out in the data cache rather than on the hard disk.

o

Log entries are not written directly to the log volume, but first into a buffer (LOG_IO_QUEUE) so as to be able to write several log entries together asynchronously with one I/O on the hard disk. Only once a LOG_IO_QUEUE page is full is it written to the hard disk. However, there are situations in which you need to write LOG entries from the LOG_QUEUE onto the hard disk as quickly as possible, for example if transactions are completed (COMMIT,ROLLBACK). The transactions wait until the log writer reports the OK informing them that the log entry is on the hard disk. Firstly, this means that is important to use the quickest possible hard disks for the log volume(s) and secondly, you must ensure that no LOG_QUEUE_OVERFLOWS occur in production operation. If the LOG_IO_QUEUE is full, then all transactions that want to write LOG entries must wait until free memory becomes available again in the LOG_IO_QUEUE. At LOG_QUEUE_OVERFOLWS (Transaction DB50 -> Current Status -> Activities Overview -> LOG_IO_QUEUE Overflow) you need to increase the LOG_IO_QUEUE parameter.

98

SAP NETWEAVER 7.1 AND HIGHER

o

PERFORMANCE CHECK GUIDE

In accordance with Note 819324 - FAQ: MaxDB SQL optimization, check which SQL statements are responsible for the most disk accesses, and whether they can be optimized. You should also optimize SQL statements that have a poor runtime and poor selectivity.

Bottleneck Analysis The Bottleneck Analysis function is available in transaction ST04 under the Problem Analyzes  Performance  Database Analyzer  Bottlenecks. You can use this function to activate the corresponding analysis tool dbanalyzer. The dbanalyzer then collects important performance data every 15 minutes and evaluates this for possible bottlenecks.

The result of the bottleneck analysis is output in text format to provide a quick overview of possible causes of performance problems. For a more detailed description of the bottleneck messages, see the online documentation (http://help.sap.com) in the SAP Web Application Server area and search for "SAP DB bottleneck analysis messages". The dbanalyzer tool can also be started at operating system level. It logs its analysis results in files with date stamps in the subdirectory \analyzer (you can find the actual directory by double-clicking on ‘Properties’ in DB50) of the relevant database instance.

9.4

Monitoring Database Tables

Some of the tables of SAP PI grow very quickly and can cause severe performance problems if archiving or deletion does not take place frequently. For troubleshooting, see SAP Note 872388 – Troubleshooting Archiving and Deletion in PI. Procedure Log on to your Integration Server and call transaction SE16. Enter the table names as listed below and execute. In the following screen simply press the button “Number of Entries”. The most important tables are:  SXMSPMAST (cleaned up by XML message archiving/deletion). If you use the switch procedure, you have to check SXMSPMAST2 as well.  SXMSCLUR / SXMSCLUP (cleaned up by XML message archiving/deletion). If you use the switch procedure, you have to check SXMSPCLUR2 and SXMSCLUP2 as well.

99

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

 SXMSPHIST (cleaned up by the deletion of history entries). If you use the switch procedure, you have to check SXMSPHIST2 as well.  SXMSPFRAWH and SXMSPFRAWD (cleaned up by the performance jobs, see SAP Note 820622 – Standard Jobs for XI Performance Monitoring).  SWFRXI* (cleaned up by specific jobs, see SAP Note 874708 - BPE HT: Deleting Message Persistence Data in SWFRXI*)  SWWWIHEAD (cleaned up by work item archiving/deletion).

 Use an appropriate database tool, for example, SQLPLUS for Oracle for the database tables of the J2EE schema. Main tables that could be affected of growth:  PI Message tables: BC_MSG and BC_MSG_AUDIT (when audit log persistence is enabled).  PI message logging information when staging/loging is used: BC_MSG_LOG_VERSION  XI IDoc tables (if IDoc persistence is activated): XI_IDOC_IN_MSG and XI_IDCO_OUT_MSG  Check for all tables if the number of entries is reasonably small or remains roughly constant over a period of time. If that is not the case, check your archiving/deletion setup.

100

SAP NETWEAVER 7.1 AND HIGHER

10

PERFORMANCE CHECK GUIDE

TRACES, LOGS, AND MONITORING DECREASING PERFORMANCE

10.1 Integration Engine There are three important locations for the configuration of tracing logging in the Integration Engine: The pipeline settings of PI itself, the Internet Communication Manager (ICM), and the gateway. All three are heavily involved in message processing and therefore their tracing or logging settings can have an impact on performance. On top of this, you might have changed the SM59 RFC destinations and have switched on the trace in order to analyze a problem: If so, then this needs to be reset as well. Procedure  Start transaction SXMB_ADM and navigate to Integration Engine Configuration  Specific Configuration and search for the following entries:

Category

Parameter

Subparameter

Current Value

Default

RUNTIME

TRACE_LEVEL





1

RUNTIME

LOGGING





0

RUNTIME

LOGGING_SYNC





0

Set the above parameters back to the default value which is the recommended value by SAP.  Start transaction SMICM and check the value of the Trace Level that is displayed in the overview (third line). Set the trace level to ‘1’ if not already the case. You can do so by navigating to Goto  Trace Level  Set  Start transaction SMGW, navigate to Goto  Parameters  Display and check the parameter Trace Level. Set the trace level to ‘1’ if not already the case. You can do so by navigating to Goto  Trace  Gateway  Reduce Level (or Increase Level, respectively).  Start transaction SM59 and search the following RFC destination for the flag “Trace” on the tab Special Options and remove the Flag “Trace” if it has been set:

   

AI_RUNTIME_JCOSERVER AI_DIRECTORY_JCOSERVER LCRSAPRFC SAPSLDAPI It is possible that other RFC destinations are used for sending out IDocs and similar

101

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

10.2 Business Process Engine You only have to check the Event Trace in the Business Process Engine. Procedure  Call transaction SWELS and check if the Event Trace is switched on. The recommended setting is ‘off’ (as shown in the screenshot below).

10.3 Adapter Framework Since the Adapter Framework runs on the J2EE Engine, the tracing and logging can be conveniently checked and controlled using NetWeaver Administrator. To improve the performance SAP recommends that you set all trace levels to default. Higher trace levels are only acceptable in productive usage during problem analysis. Below you will find a description on how to set the default trace level for all locations at once. It is of possible to do it for every location separately, but this way you might forget to reset one of the locations. Procedure  Start the NetWeaver Administrator and navigate to Problem Management (or Troubleshooting in NW 7.3 and higher)  Logs and Traces and choose Log Configuration. In the Show drop down list box choose Tracing Locations. Select the Root Location in the tree and press the button Default Configuration. This will set the default configuration for all locations.

 Start the NetWeaver Administrator and navigate to Troubleshooting  Logviewer (direct link /nwa/links). Open the View “Developer Trace” and check if you have very frequent, reoccurring exceptions that fill up the trace. Analyze the exception and check what is causing this (e.g. a wrong configuration of a Communication Channel causing repeated traces. If you see very frequent errors for which you cannot find the root cause open a customer incident at SAP.

102

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

 An aggregated view of Java exceptions (by number, date, instance, server node, etc) are reported and available in SAP Solution Manager, Root Cause Analysis, Exception Analysis functionality:

10.3.1 Persistence of Audit Log information in PI 7.10 and higher With SAP NetWeaver PI 7.1 the audit log is not persisted for successful messages in the database by default to avoid performance overhead. Therefore, the audit log is only available in the cache for a limited period of time (based on the overall message volume). Audit log persistence can be activated temporarily for performance troubleshooting where audit log information is required. To do so, use the NWA and change the parameter “messaging.auditLog.memoryCache” to false in service XPI Service: Messaging System by setting the parameter to false. Details can be found in SAP Note 1314974 - PI 7.1 AF Messaging System audit log persistence. Only do so temporarily during the time of troubleshooting to avoid any performance problems from the additional persistence. After implementing Note 1611347 - New data columns in Message Monitoring additional information like message processing time in ms and server node is visible in the message monitor.

103

SAP NETWEAVER 7.1 AND HIGHER

11

PERFORMANCE CHECK GUIDE

ERRORS AS SOURCE OF PERFORMANCE PROBLEMS

If errors occur within messages or within technical components of the SAP Process Integration, then this can have a severe impact on the overall performance. Thus, to solve performance problems, it is sometimes necessary to analyze the log files of technical components and to search for messages with errors. To search for messages with errors, use the PI Admin Check which is available using SAP Note 884865 - PI/XI Admin Check. Procedure  ICM (Internet Communication Manager): Start transaction SMICM, open the log file (Shift + F5 or GoTo Trace File  Show All), and check for errors.  Gateway: Start transaction SMGW, open the Log File (CTRL + Shift + F10 or Goto  Trace  Gateway  Display File), and check for errors.  System Log: Start transaction SM21, choose an appropriate time interval, and search the log entries for errors. Repeat the procedure for remote systems if you are using dialog instances.  ABAP Runtime Errors: Start transaction ST22, choose an appropriate time interval, and search for PI-related dumps. In general the number of ABAP dumps should be very small on a PI system and therefore all occurring dumps should be analyzed.  Alerts / CCMS: Start transaction RZ20 and search for recent alerts.  Work Process and RFC trace: In the PI work directory check all files which begin with dev_rfc* or dev_w* for errors.  J2EE Engine: In the PI work directory search for errors in file dev_server0.out. If you are using more than one J2EE server node, check dev_server.out as well.  Applications running on the J2EE Engine: The traces for application-related errors are the default.trc files located in /j2ee/cluster/server/log directory. To monitor exceptions across multiple server nodes, the best tool to use is the LogViewer service of the NWA. This is available via Problem Management  Logs and Traces and choose Log Viewer. Search for exceptions in the area of the performance problem, for example, the PI Messaging system.

104

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

APPENDIX A A.1

Wily Introscope Transaction Trace

As mentioned above, the Wily Transaction trace can be used to trace expensive steps that were noticed during the mapping or module processing. The transaction trace will allow you to drill down further into Java performance problems and to distinguish if it is a pure coding problem or caused by a look-up to a remote system or a slow connection to the local database. The Wily Transaction trace can be used to trace all steps that exceed a specific duration on the Java stack. If a user is known (for example, for SAP Enterprise Portal) the tracing can be restricted to a specific user. In PI, if you encounter a module that lasts several seconds then you can restrict the tracing as shown below. Note: Starting the Transaction Trace increases the data collection on the satellite system (PI) and is therefore only recommended in productive environment for troubleshooting purposes.

With the selection above the trace will run for 10 minutes (this is the maximum and it can be canceled earlier) and will trace all steps exceeding 3 seconds for the agents of the PI system. If such long-running steps are found then a new window will be displayed listing these steps. In the screenshot below you can see the result in the Trace View. The Trace View shows the elapsed time from left to right– in the example below around 4.8 seconds. From top to bottom we can see the call stack of the thread. In general we are interested in long-running threads on the bottom of the trace view. A longrunning block at the bottom means that this is the lowest level coding that was instrumented and which is consuming all the time. In the example below we see a mapping call that is performing many individual database statements – this will become visible by highlighting the lowest level. In such a case you have to review the coding of the mapping to see if the high amount of database calls can be summarized in one call.

105

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

Another case that is often seen is that a lookup using JDBC to a remote database or RFC to an ABAP system takes a long time in the mapping or adapter module. In such a case there will be one long block at the bottom of the transaction trace that also gives you some details about the statement that was executed. A.2

XPI inspector for troubleshooting and performance analysis

The XPI Inspector is an additional tool that provides enhanced logging and tracing capabilities with a main focus on troubleshooting. But the tool can also be used for troubleshooting of performance issues. General information about the tool can be found at Note 1514898 - XPI Inspector for troubleshooting PI. The tool can be called using the following URL on your system: http://host:java-port/xpi_inspector. To analyze performance problems typically the example “51 – “Performance Problem)” is used. As a basic measurement it allows you to take multiple thread dumps in specific time intervals. These thread dumps can be analyzed later by your administrators or SAP support to identify what is causing the problem. The Thread Dump Viewer tool (TDV) from Note 1020246 - Thread Dump Viewer for SAP Java Engine can be used to analyze the produced thread dumps (inside the results archive that can be downloaded to your local PC).

106

SAP NETWEAVER 7.1 AND HIGHER

PERFORMANCE CHECK GUIDE

In addition the tool allows you to do JVM profiling be either doing JVM Performance tracing or JVM Memory Allocation tracing. This can help to understand in detail which steps of the processing are taking a long time. As output the tool provides *.prf files that can be loaded into the JVM Profiler. More information can be found at http://wiki.scn.sap.com/wiki/display/ASJAVA/Java+Profiling. Please Note that these traces (especially the memory profiling) will cause a high overhead on the J2EE engine and are therefore not recommended to be used in production. Instead it is advisable to reproduce the problem on your QA system and use the tool there.

107

www.sap.com

Document Version 3.0 March 2014

www.sap.com

© 2014 SAP AG. All rights reserved. SAP, R/3, SAP NetWeaver, Duet, PartnerEdge, ByDesign, SAP BusinessObjects Explorer, StreamWork, SAP HANA, and other SAP products and services mentioned herein as well as their respective logos are trademarks or registered trademarks of SAP AG in Germany and other countries. Business Objects and the Business Objects logo, BusinessObjects, Crystal Reports, Crystal Decisions, Web Intelligence, Xcelsius, and other Business Objects products and services mentioned herein as well as their respective logos are trademarks or registered trademarks of Business Objects Software Ltd. Business Objects is an SAP company. Sybase and Adaptive Server, iAnywhere, Sybase 365, SQL Anywhere, and other Sybase products and services mentioned herein as well as their respective logos are trademarks or registered trademarks of Sybase Inc. Sybase is an SAP company. Crossgate, m@gic EDDY, B2B 360°, and B2B 360° Services are registered trademarks of Crossgate AG in Germany and other countries. Crossgate is an SAP company. All other product and service names mentioned are the trademarks of their respective companies. Data contained in this document serves informational purposes only. National product specifications may vary. These materials are subject to change without notice. These materials are provided by SAP AG and its affiliated companies ("SAP Group") for informational purposes only, without representation or warranty of any kind, and SAP Group shall not be liable for errors or omissions with respect to the materials. The only warranties for SAP Group products and services are those that are set forth in the express warranty statements accompanying such products and services, if any. Nothing herein should be construed as constituting an additional warranty.