A white paper prepared with Oracle and IBM technical collaboration March 2010

Oracle Real Application Clusters on IBM AIX Best practices in memory tuning and configuring for system stability

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

Executive Overview ........................................................................... 2 Introduction ....................................................................................... 2 Problem validation ......................................................................... 3 Examining the AIX Error Logging................................................... 3 Examining the Oracle oprocd logging ............................................ 4 Recommendations for system stability............................................... 5 1. Implement AIX tuning recommendations for Oracle ................... 5 2. Modify Oracle ‘diagwait’ parameter............................................ 6 3. Install the required updates and patches ................................... 8 4. Reduce heavy paging activity. ................................................. 11 Appendix A: Oprocd logging examples ............................................ 14 Clean oprocd log files .................................................................. 14 Oprocd log files showing scheduling delays................................. 14 Oprocd log files after a node reboot or CRS restart ..................... 14 Appendix B: Example script for setting the correct VMM settings .... 16 References...................................................................................... 17

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

Executive Overview
IBM and Oracle have a long standing history of technical collaboration to enhance Oracle products on IBM Power™ Systems. This joint commitment to support the Power System platform’s performance, scalability and clustering capabilities allows customers to effectively deploy Oracle Real Application Clusters (RAC) on a broad range of configurations. To support customers who choose to deploy Oracle RAC on IBM Power Systems servers IBM and Oracle have prepared this recommendation of best practices for managing memory use on AIX® systems running Oracle RAC. Implementing these recommendations will provide customers with the best possible availability and performance of the Oracle Database, and reduce the potential for node evictions due to memory over commitment. Key elements of the best practices are tuning AIX and Oracle RAC, moving to the current Oracle Database 10g Release2 patch set and monitoring the system resources to insure memory is not over committed.

Introduction
Customers who experience Oracle Real Application Clusters (RAC) node evictions due to excessive AIX kernel paging should carefully review and implement these recommended best practices. Testing and experience have found that memory over commitments may cause scheduling delays for Oracle’s ‘oprocd’ process resulting in node evictions. Implementing all of these recommendations will reduce scheduling delays and corresponding oprocd initiated evictions.

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

Problem validation
This paper addresses the best practices for environments experience node evictions caused by critical processes not being able to get scheduled in a timely fashion on AIX due to memory over commitment. To validate that node evections are caused by this situation, the following validation steps should be taken.

Examining the AIX Error Logging
When an Oracle RAC cluster node is rebooted for cluster integrity this can be done by several cluster processes. When a node gets rebooted due do scheduling problems, the process, which would initiate the reboot, is oprocd. The first step in validating that scheduling delays caused a node reboot is to confirm if the oprocd process rebooted the node. When the oprocd process reboots the node there should be only one entry in the output of ‘errpt –a’ created at the time of the reboot. The error created in the AIX Error Logging subsystem should look similar to the following example:

LABEL: IDENTIFIER:

REBOOT_ID 2BFA76F6

Date/Time:

Thu Apr 23 16:00:58 PDT 2009

Sequence Number: 10 Machine Id: Node Id: Class: Type: 00C47AAC4C00 racha908 S TEMP

Resource Name: Description

SYSPROC

SYSTEM SHUTDOWN BY USER Probable Causes SYSTEM SHUTDOWN

Detail Data USER ID 0 0=SOFT IPL 1=HALT 2=TIME REBOOT 0 TIME TO REBOOT (FOR TIMED REBOOT ONLY) 0

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

There should not be a ‘SYSDUMP’ entry in the ‘errpt –a’ around the time of the reboot since ‘oprocd’ does not initiate a ‘SYSDUMP’. A ‘SYSDUMP’ entry is an indication that other problems may be the root cause of node reboots.

Examining the Oracle oprocd logging
Please note that to be able to analyze the oprocd log files to identify reboots due to AIX scheduling delays, all required Oracle patches should be in place (see Recommendation #3 below). When the required patches are in place, a clear message in the most recent /etc/oracle/oprocd/.oprocd.lgl. file should also show that oprocd rebooted the node.

Apr 23 16:00:04.327787 | LASTGASP | AlarmHandler: interval(1000 msec)+margin(10000 msec).

timeout(11917 msec) exceeds

Rebooting NOW.

Oracle Support Data - Previous intervals (LIFO): 999ms 999ms 1000ms

>

In this example the oprocd margin was already increased from 500 ms to 10 seconds, according to recommendation #2 in this paper. This example shows a system being rebooted due to a scheduling delay of 11.9 seconds. This delay exceeded the sum of the oprocd intervals (1 second) and the oprocd margin (10 seconds). Therefore the reboot was initiated.

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

Recommendations for system stability
The following recommendations will help ensure AIX systems running Oracle RAC will run in a stable and reliable manner. Unless otherwise indicated all recommendations apply to AIX 5.2, AIX 5.3 and AIX 6.1.

1. Implement AIX tuning recommendations for Oracle
The first step in removing scheduling delays due to AIX kernel paging is to make sure a system has the correct AIX vmo parameters set. The following parameters should be verified and set on all RAC nodes. vmo -p -o maxperm%=90; vmo -p -o minperm%=3; vmo -p -o maxclient%=90; vmo -p -o strict_maxperm=0; vmo -p -o strict_maxclient=1; vmo -p -o lru_file_repage=0; vmo -r -o page_steal_method=1; chdev –l sys0 –a ‘minpout=4096 maxpout=8193’; (Note: 8193 is the correct value here, although, ‘8192” might look more reasonable.) Appendix B provides an example script to set these parameters. Enable write/commit behind for remote JFS and JFS2 file systems that are NFS mounted on a RAC node and used for backups. This allows AIX to detect when a file is being written serially, and so the modified pages can be flushed to the paging device in smaller chunks rather than (potentially) all at one time,. This will slow down individual write operations slightly (for that specific file system) but has a system performance benefit similar to I/O pacing. For NFS file systems, the file system should be mounted with the options shown in the following example. mount –o combehind,numclust=128 : The “mount” command can then be used to confirm that the options are in use for the file system.

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

Users should also review all the tuning recommendations in the document “Tuning IBM AIX 5.3 and AIX 6.1 for Oracle Database” which is available at the following location: http://www.ibm.com/servers/enable/site/peducation/wp/162b6/162b6.pdf

2. Modify Oracle ‘diagwait’ parameter.
The oprocd process runs with two important parameters: interval and margin. The interval parameter is the time that the process sleeps before trying to get scheduled again. The default interval is one second. The margin parameter is the time, which oprocd is allowed to deviate from the scheduling interval time. The default is 500 milliseconds. To make the oprocd process less sensitive to scheduling delays, the margin can be increased to 10 seconds. Changing the Clusterware parameter diagwait to 13 is the Oracle supported technique to change the oprocd margin to 10 seconds. Please note that 13 is the only allowed value for setting the diagwait parameter to. Any value other than 13 (or unset) is not allowed and not supported.
Procedure for changing the Clusterware parameter diagwait

1.

Stop Oracle Clusterware on ALL cluster nodes by executing the following command on each node as the root user.

#crsctl stop crs #/bin/oprocd stop

2.

Ensure that Clusterware stack is down by running the “ps” command. Executing this command should return no processes on any of the cluster nodes.

#ps -ef |egrep "crsd.bin|ocssd.bin|evmd.bin|oprocd"

3.

From one node of the cluster, change the value of the "diagwait" parameter to 13 seconds by issuing the command as root:

#crsctl set css diagwait 13 -force

4.

Confirm that diagwait is set successfully by executing the following command. The command should return 13. If diagwait is not set, the following message will be returned "Configuration parameter diagwait is not defined"

#crsctl get css diagwait

5.

Restart the Oracle Clusterware by running the following command on all the nodes:

#crsctl start crs

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

6.

Validate that the node is running by executing:

#crsctl check crs

An example of setting diagwait to 13: oratest@racha906 /home/oratest > crsctl set css diagwait 13 Configuration parameter diagwait is now set to 13.

oratest@racha906 /home/oratest > crsctl get css diagwait 13oratest@racha906 /home/oratest >

Note that the new value for ‘diagwait’ is 13 but printed with no carriage return on the line.

When we startup the Clusterware stacks again after we change the diagwait to 13, we will see the following log lines in the oprocd log file /etc/oracle/oprocd/.oprocd.log:
Apr 23 16:04:34.413 | INF | monitoring started with timeout(1000), margin(10000), skewTimeout(250) Apr 23 16:04:34.495 | INF | fatal mode startup, setting process to fatal mode Apr 23 16:04:40.385 | INF | enabling fatal mode as per client request

Note that now the margin is set to 10000. For more information on this topic refer to Oracle’s MetaLink Docid number 559365.1 “Using Diagwait as a diagnostic to get more information for diagnosing Oracle Clusterware Node evictions.”

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

3. Install the required updates and patches
To further reduce the risk of memory over commitment causing delays in process scheduling causing evictions the following Oracle and AIX software versions should be used.
Required patches and updates for Oracle RAC:

Install the correct patch sets and recommended patch bundles for Oracle Clusterware and Oracle RAC. The following is a list of the minimal required Patch Set levels; higher Patch Set levels can be used instead. • For Oracle 10g Release 2 o o o o 10g Release 2 (10.2.0.4) Patch Set 3 Recommended RAC Bundle 3 Recommended CRS Bundle 3 Patch: Patch: Patch: 6810189 8344348 7715304

If running AIX 5.2 ML07 or a higher ML then also install the Oracle patch for Bug 7321562 • Note, AIX 5.2 entered extended support in April 2009, customers are encouraged to upgrade to AIX 5.3 or AIX 6.1.

•

For Oracle 11g Release 1 o 11g Release 1 (11.1.0.7) Patch Set 1 Patch: 6890831

These patch bundles contain fixes that will: Insure that all ‘oprocd’ messages are being logged and saved in the /etc/oracle/oprocd/*log* and /etc/oracle/oprocd/*lgl* files. Allow histogram data of ‘oprocd’ delays to be collected in /etc/oracle/oprocd/*log* to help in evaluating the severity of the ‘oprocd’ delay problem. Insure that CRS and ‘oprocd’ are running with the correct priority, scheduling policy, and pinning memory correctly.

Required patch set and updates for AIX:

•

For AIX 6.1 o Follow the AIX requirements in the Oracle Release notes located in the Oracle Database Documentation Library.

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

• • • For AIX 5.3 o •

Oracle 10gR2 release notes are B19074-11 – or higher Oracle 11gR1 release notes are B32075-07 – or higher

AIX 5.3 ML06 - or higher ML, in addition to the AIX requirements in the Oracle Release notes.

For AIX 5.2 o Follow the AIX requirements in the Oracle Release notes.

In general, please review the following My Oracle Support Notes for the latest patch updates: • • DocID 756671.1 – Oracle Recommended Patches -- Oracle Database DocID 811293.1 – RAC Assurance Support Team: RAC Starter Kit and Best Practices (AIX)

Validating oprocd and cssd process priority, scheduling policy and memory pinning.

After installation of the required Oracle patches and AIX patches: oprocd.bin will have a priority (PRI) of ‘1 or 0’ and a scheduling policy (SCH) of ‘2’. Note, the PRI value for oprocd will depend on the Oracle version and patches applied: for CRS 10.2.0.4 before bundle 6 is applied: 1 for CRS 10.2.0.4 after bundle 6 is applied: 0 for CRS 11gR1: 0 ocssd.bin should have a priority (PRI) of ‘0’ and a scheduling policy (SCH) of ‘-‘. All ‘ora_lms’ processes should have a priority (PRI) of ‘39’ and a scheduling policy (SCH) of ‘--‘.

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

Priority and scheduling policy can be verified by running the following command:
> ps -ef -o pid,pri,sched,nice,args |egrep 'oprocd|ora_lms|ocssd|COMMAND' |grep

-v grep |sort +4 PID PRI SCH NI COMMAND 503810 40 0 0 /bin/sh -c ulimit -c unlimited; cd || exit

/oratest/Vtest/CRS905/log/racha905/cssd; /oratest/Vtest/CRS905/bin/ocssd $? 974890 602306 962684 60 0 0 0 20 /bin/sh /etc/init.cssd oprocd - -- /oratest/Vtest/CRS905/bin/ocssd.bin

2 -- /oratest/Vtest/CRS905/bin/oprocd.bin run -t 1000 -m 10000 -hsi

5:10:50:75:90 –f 622646 700516 577726 688216 827452 585824 1056948 422014 39 39 39 39 39 39 39 39 2 -- ora_lms0_swing1 2 -- ora_lms0_tpch5 2 -- ora_lms1_swing1 2 -- ora_lms1_tpch5 2 -- ora_lms2_swing1 2 -- ora_lms2_tpch5 2 -- ora_lms3_swing1 2 -- ora_lms3_tpch5

Also note that when all patches are in place, the process names will have changed. $CRS_HOME/bin/oprocd 10000 -hsi 5:10:50:75:90 –f $CRS_HOME/bin/ocssd $CRS_HOME/bin/oprocd.bin run -t 1000 -m $CRS_HOME/bin/ocssd.bin

In addition oprocd.bin and ocssd.bin should have the majority of their memory pinned. This should be verified using the following command. The “Pin” value should be close to the size of the “Inuse” value.
> ps -elf -o "pid,args"|egrep "oprocd|ocssd.bin" |grep -v grep |awk '{ print $1 }' |xargs svmon -P |egrep "oprocd|ocssd.bin|Command Pid Command 868420 ocssd.bin Pid Command 294970 oprocd.bin Pid Command Inuse 91412 Inuse 81510 Inuse Pin 82879 Pin 73242 Pin Pgsp 0 Pgsp 0 Pgsp Virtual 64-bit Mthrd 91114 Y Y 16MB N 16MB N 16MB

Virtual 64-bit Mthrd 81487 Y N

Virtual 64-bit Mthrd

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

4. Reduce heavy paging activity.
As with most operating systems, heavy paging on an AIX system can cause scheduling delays. Very heavy paging can cause longer scheduling delays, which can interfere with the critical Oracle processes like oprocd and cssd. To prevent delays due to heavy paging, the system should be monitored and tuned to avoid heavy paging. If a system is seeing heavy paging, there are two ways to avoid heavy paging: Tune the workload to reduce its memory usage Increase the amount of physical memory allocated to the workload Note: Increasing paging space without any increase in physical memory will not help reduce paging activity.

Monitoring memory usage and paging

Monitoring a workload is key to identifying why a system is heavily paging and to prevent performance impacts to the workload. Memory utilization can be monitored using many tools and management suites, but some basic monitoring can be achieved using the AIX tools vmstat and svmon. The vmstat tool can be used to monitor the amount of memory being used as well as the rate of paging on the system. The key metrics to observe are active virtual memory (avm), page-in rate (pi), and page-out rate (po). The avm field will report the total amount of virtual memory in-use in units of 4K pages. When the amount of virtual memory on the system gets close to the memory size of the LPAR, this can be treated as an early indication that the system is getting close to running out of physical memory and may start paging. It is important to note that the number of free pages (fre) is not a good indication of whether a system is low on free memory. AIX aggressively uses memory to cache file data, and thus, it is not unusual to see a low number of free pages even when the amount of active virtual memory is low. A better indication of whether a system is close to paging is to look at the active virtual memory (avm) field and compare it to the total amount of memory. Once the amount of active virtual memory (avm) exceeds the amount of memory in the system, AIX will begin paging. The page-in (pi) and page-out (po) fields can be used to monitor the paging activity. These fields report the rate of page-in and page-out operations on a system. In the following vmstat example, the small amount of free memory (fre) leads to high paging rates for both page in (pi) and page out (po).

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

The following is an example of a system incurring heavy paging that resulted in an ‘oprocd’ eviction. These high levels of paging should be avoided.

> vmstat –t 2 kthr time ----- ----------- ------------------------ ------------ ----------------------- ------r b avm fre re pi po fr sr cy in sy cs us sy id wa pc ec memory page faults cpu

hr mi se 2 17 8056132 12319 15:11:28 1 16 8056912 12330 15:11:30 2 18 8057768 12568 15:11:32 1 17 8067254 12317 54.3 15:11:34 1 18 8072089 12492 52.2 15:11:36 1 16 8084793 12577 55.2 15:11:38 1 12 8105868 13164 63.2 15:11:40 1 16 8106267 18276 15:11:42 1 16 8106273 20478 15:11:47 1 9 8107224 20111 0 179 491 515 515 0 1377 13155 7880 2 52 6 39 1.21 0 5 446 0 0 0 422 48 221 0 46 16 38 1.01 50.5 0 4 2766 0 0 0 492 56 191 0 46 16 37 1.03 51.4 0 0 10711 10224 141211 0 823 31 1807 3 54 1 42 1.26 0 0 6393 6336 6355 0 616 27 1156 2 48 0 50 1.10 0 0 2504 2432 2436 0 434 52 691 1 46 0 53 1.04 0 0 4617 4864 4878 0 538 149 1203 2 48 0 51 1.09 0 0 548 320 320 0 390 36 571 0 46 0 54 1.01 50.7 0 0 396 320 320 0 379 39 429 0 46 0 54 1.01 50.7 0 0 386 512 512 0 395 34 534 0 45 0 54 1.01 50.5

60.5 15:11:49

Another tool that can be used to monitor memory utilization of a system is svmon. Starting with AIX 5.3 TL09 and AIX 6.1 TL02 svmon now reports an “available” metric. This metric can be used to more easily determine how much remaining memory is available to applications. The available metric reports the amount additional amount of physical memory that can be used for applications without incurring paging. When the amount of available memory gets low, this is an indication that the system is close to paging.

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

This can be seen in the following example: root@racha905 / > svmon -G -O unit=auto ------------------------------------------------------------------------------size memory pg space 27.2G 24.5G inuse 8.04G 24.4M free 19.2G pin 2.45G virtual 7.43G available 19.2G

work pin in use 1.26G 7.43G

pers 4K 311.66M

clnt 388K 311.68M

other 1.19G

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

Appendix A: Oprocd logging examples
Clean oprocd log files
For /etc/oracle/oprocd/.oprocd.log
Apr 23 16:04:34.413 | INF | monitoring started with timeout(1000), margin(10000), skewTimeout(250) Apr 23 16:04:34.495 | INF | fatal mode startup, setting process to fatal mode Apr 23 16:04:40.385 | INF | enabling fatal mode as per client request

For /etc/oracle/oprocd/.oprocd.lgl
Apr 23 16:04:34.405420 | LASTGASP | InitLastGasp: gasp file Initial write/allocate for last

Oprocd log files showing scheduling delays
For /etc/oracle/oprocd/.oprocd.log
Apr 23 16:04:34.413 | INF | monitoring started with timeout(1000), margin(10000), skewTimeout(250) Apr 23 16:04:34.495 | INF | fatal mode startup, setting process to fatal mode Apr 23 16:04:40.385 | INF | enabling fatal mode as per client request

For /etc/oracle/oprocd/.oprocd.lgl
Apr 23 16:04:34.405420 | LASTGASP | InitLastGasp: gasp file Apr 24 05:20:03.665 | INF | TrackHistoricalTrends: in 10 to 50 percentile Apr 26 05:26:23.593 | INF | TrackHistoricalTrends: in 10 to 50 percentile added first sample 2642327278 added first sample 3242554777 Initial write/allocate for last

Note that 2 entries have been recorded where the delay in scheduling was approximately 3.24 and 2.64 seconds respectively which are greater than the old 1.5 second limit but less than the new 11 second limit.

Oprocd log files after a node reboot or CRS restart
During CRS restart, the previous files will be renamed with a time stamp appended to the file names:

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

For /etc/oracle/oprocd/.oprocd.log.
Apr 23 16:04:34.405420 | LASTGASP | InitLastGasp: gasp file Apr 24 05:20:03.665 | INF | TrackHistoricalTrends: in 10 to 50 percentile Apr 26 05:26:23.593 | INF | TrackHistoricalTrends: in 10 to 50 percentile added first sample 2642327278 added first sample 3242554777 Initial write/allocate for last

For /etc/oracle/oprocd/.oprocd.lgl.
Apr 28 16:00:04.327787 | LASTGASP | AlarmHandler: interval(1000 msec)+margin(10000 msec). timeout(11917 msec) exceeds

Rebooting NOW.

Oracle Support Data - Previous intervals (LIFO): 999ms 999ms 1000ms >

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

Appendix B: Example script for setting the correct VMM settings
The following example script can be used to set the correct VMM parameters according to the recommendations in this white paper:
#!/usr/bin/ksh

vmo -p -o maxperm%=90; vmo -p -o minperm%=3; vmo -p -o maxclient%=90; vmo -p -o strict_maxperm=0; vmo -p -o strict_maxclient=1; vmo -p -o lru_file_repage=0; vmo -r -o page_steal_method=1; chdev –l sys0 –a ‘minpout=4096 maxpout=8193’;

Oracle White Paper — Oracle Real Application Clusters on IBM AIX - Best practices for system stability

References
My Oracle Support DocID # 265769.1 “Troubleshooting CRS Reboots” My Oracle Support DocID # 419312.1 “Node reboots due to oprocd on AIX” My Oracle Support DocID # 559365.1 “Using Diagwait as a diagnostic to get more information for diagnosing Oracle Clusterware Node evictions”

Oracle Real Application Clusters on IBM AIX – Best practices in memory tuning and configuring for system stability - Version: 1.2, March 2010 Authors: Rick Piasecki (IBM), Daniel Dibbets, Wayne Martin (IBM), Rahim Mau (Oracle) Contributing Authors: David Hepkin(IBM), Anil Nair (Oracle), Dennis Massanari (IBM), Markus Michalewicz (Oracle) Oracle Corporation World Headquarters 500 Oracle Parkway Redwood Shores, CA 94065 U.S.A. Worldwide Inquiries: Phone: +1.650.506.7000 Fax: +1.650.506.7200 oracle.com 0109 Copyright © 2009, Oracle and/or its affiliates. All rights reserved. This document is provided for information purposes only and the contents hereof are subject to change without notice. This document is not warranted to be error-free, nor subject to any other warranties or conditions, whether expressed orally or implied in law, including implied warranties and conditions of merchantability or fitness for a particular purpose. We specifically disclaim any liability with respect to this document and no contractual obligations are formed either directly or indirectly by this document. This document may not be reproduced or transmitted in any form or by any means, electronic or mechanical, for any purpose, without our prior written permission.

Oracle is a registered trademark of Oracle Corporation and/or its affiliates. Other names may be trademarks of their respective owners.