1. Presentation 36492 OracleAS Tuning Techniques Donald K. Burleson Col. John Garmany Burleson Oracle Consulting

4. Oracle Training by Don Burleson

5. www.rampant.cc Oracle Books from $9.95

6. Guidehorse.com

7. John Garmany Retired Army Lt. Col. - West Point Graduate - Masters Degree Information Systems - Graduate Certificate in Software Engineering - Airborne Ranger

8.  On-site custom Oracle training  Oracle Tuning & Oracle Support  Remote DBA Support

9. Topics:  OracleAS Tuning Approach  OracleAS Monitoring  Tuning with RAM  Load Balancing

10. Enhancement request for OracleAS:  Add a new global parameter: run_fast = yes

11. OracleAS Architecture Web Cache Web Cache Web Cache Web Cache HTTP Server HTTP Server HTTP Server HTTP Server HTTP Server HTTP Server Database Files RAC Server RAC Server RAC Server RAC Server RAC Server Internet

12. Keys to Success  Parameter Tuning  RAM Cache Tuning  Server Tuning

13. Parameter Tuning  OracleAS parameters - Adjusting the Oracle9iAS configuration parameters for each Oracle9iAS component has influence performance and throughput.  Database parameters – Because most Oracle9iAS systems are disk I/O intensive, adjusting the Oracle database parameters for the Infrastructure database (iasdb) and the back-end database can heavily influence performance.

14. RAM Tuning  Data buffer tuning – Adding RAM to the database db_cache_size on the Oracle Infrastructure and back- end database can greatly reduce disk I/O and improve throughput.  Web cache tuning – Adding RAM to the Oracle9iAS web cache can improve the delivery rates of HTML and XML though the Oracle HTTP Server (OHS).

15. Server tuning  Hardware configuration – Adding RAM of CPU resources to existing servers will improve the throughput on the server  Hardware load balancing – The addition of new servers to the Oracle9iAS farm and relocating Oracle9iAS components. Spare servers can be configured with both Web Cache and App Server, and the appropriate components can be started as- needed.  Server parameter tuning – Adjusting the parameters on your server can have a huge impact on the performance of the OracleAS.

16. Monitoring Techniques  Response Time Monitoring – DCM and OEM  Wait Event Monitoring – Determine the source of Latency for each Component.  Server Resources – Once the farm is tuned, overloads can be addressed with dynamic server allocation.

17. Wait Event Monitoring (for isadb and database) How would you tune this database? % Total Event Waits Time (s) Ela Time ------------------------------ ----- -------- -------- CPU time 30 71.43 db file parallel write 95 1 23.53 control file sequential read 54 1 2.33 log file parallel write 62 0.95 db file sequential read 20 0.68 How would you tune this database? % Total Event Waits Time (s) Ela Time ------------------------------ ----- -------- -------- db file sequential read 45 22 41.43 db file scattered read 95 14 25.55 control file sequential read 54 1 2.33 log file parallel write 62 0.95 db file parallel write 20 0.68

18. OracleAS Monitoring

19. Dynamic Monitoring Service (DMS)  OC4J – Measure Parse Time for Incoming Request and Free RAM in the JVM  Portal – Display Portal Metrics  Servlet – Instrument Servlets to Generate Performance Metrics  OHS – Measure Active HTTP Requests

20. DMS has over 300 metrics dmstool -l |grep completed /appsvr/OC4J:3303:6004/oc4j/default/WEBs/parseRequest.completed /appsvr/OC4J:3303:6004/oc4j/default/WEBs/processRequest.completed /appsvr/OC4J:3303:6004/oc4j/default/WEBs/resolveContext.completed /appsvr/OC4J:3303:6004/oc4j/portal/WEBs/parseRequest.completed /appsvr/OC4J:3303:6004/oc4j/portal/WEBs/processRequest.completed /appsvr/OC4J:3303:6004/oc4j/portal/WEBs/resolveContext.completed /ap/OC4J:3303:6004/oc4j/syndserver/WEBs/parseRequest.completed /ap/OC4J:3303:6004/oc4j/syndserver/WEBs/processRequest.completed /ap/OC4J:3303:6004/oc4j/syndserver/WEBs/resolveContext.completed

21. Collect 100 sets at 60 second intervals dmstool -i 60 -c 100 \ /appsvr/Apache:2534:6004/Apache/handle.completed \ /appsvr/Apache:2534:6004/Apache/request.completed \ /appsvr/Apache:2534:6004/Apache/handle.completed \ /appsvr/Apache:2534:6004/Apache/request.completed >> t1.lst Output Listing Sun Jul 13 20:19:43 MDT 2003 /appsvr/Apache:2534:6004/Apache/handle.completed 240320 ops /appsvr/Apache:2534:6004/Apache/request.completed 146504 ops /appsvr/Apache:2534:6004/Apache/connection.completed 56908 ops

22. Compute delta in spreadsheet

23. Plot with Chart Wizard

24. DMS can be scripted: #!/bin/ksh PATH=$PATH:/home/oracle/oraportal904/bin export PATH # Dump Stats for Later Analysis dmstool -dump >> dumparch.lst # Dumping OHS Stats to a File dmstool -table ohs_server >> ohs.lst

25. Sending OHS stats to a flat file: # Dumping OHS Stats to a File dmstool -table ohs_server >> ohs.lst cat ohs.lst|grep connection.active > con_active.lst cat ohs.lst|grep request.active > req_active.lst cat ohs.lst|grep busyChildren.value > busy_child.lst cat ohs.lst|grep readyChildren.value > readyChild.lst cat ohs.lst|grep numChildren.value > det.lst

26. OHS Server Output Sun Jul 13 21:01:45 MDT 2003 ---------- ohs_server ---------- busyChildren.value: 16... childStart.count: 24748 ops connection.active: 24 threads... numChildren.value: 44... readyChildren.value: 27... request.avg: 15321 usecs request.completed: 150942 ops...

27. Plotting OHS response time

28. List OHS performance metrics dmstool -table ohs_module -c 1 Name: mod_oc4j.c... decline.count: 13487 ops handle.active: 0 threads handle.avg: 3 usecs handle.completed: 13487 ops handle.maxTime: 8 usecs handle.minTime: 2 usecs handle.time: 43710 usecs Name: http_core.c... decline.count: 0 ops handle.active: 0 threads handle.avg: 0 usecs handle.completed: 0 ops handle.maxTime: 0 usecs Hard to parse The output

29. Computing real response time One of the problems with the OHS statistics is that the one-time operations will skew the overall averages in the ohs_response listings. (time – min – max) real_average = ------------------------ (completed – 2) Using the data from the previous mod_oc4j.c listing, we can compute the real response time: (43,710 – 2 – 8) real_average = ------------------------ (13,487 – 2) (43,700) real_average = ------------------ = 3.24 milliseconds (13,485)

30. Using Aggrespy

32. Web Cache Monitoring

33. OracleAS Web Cache Internet Web Cache Web Server Database Trigger Programmatic

34. Web Cache Tuning  Static and Dynamic Information  Cacheability Rules  Cache Invalidations  Multi-version HTML  Rule for Each Page Component

35. Web Cache statistics:  Requests – This shows the current, average and max transaction per second. The backlog section indicates that the Web Cache is overwhelmed and another Web Cache server should be started.  Errors – This summarized the network, site busy and particle- page errors for the Web Cache.  Misses – This section shows cacheable and non-cacheable misses along with the number of refreshes for the Web Cache.  Compression – The compression sections show the total amount of RAM saved by compression and provides a great gauge of the effectiveness of the Web Cache.

36. OracleAS Load Balancing

37. Software Load Balancing  Web Cache to OHS – Web Cache interrogates OHS statistics and routes to least loaded.  OHS to Database Listener – OHS Distributes load to multiple Listeners  Database Listener – Listeners to Multiple Dispatchers under MTS, that load balance to least loaded RAC Instance.

38. OracleAS Load Balancing Web Cache Web Server Database Web Server Web Cache Instance Application Server Tier Web Cache Tier Database Server Tier

39. Hardware Load Balancing Web Cache Web Cache Web Cache Web Cache HTTP Server HTTP Server HTTP Server HTTP Server HTTP Server HTTP Server Database Files RAC Server RAC Server RAC Server RAC Server RAC Server Internet Blade Server Rack OHS & WC Oracle RAC

40. Monitoring Servers with vmstat SAMPLE_TIME=300 while true do vmstat ${SAMPLE_TIME} 2 > /tmp/msg$$ # run vmstat and direct the output into the Oracle table... cat /tmp/msg$$|sed 1,3d | awk '{ printf("%s %s %s %s %s %s\n", $1, $8, $9, 14, $15, $16) }' | while read RUNQUE PAGE_IN PAGE_OUT USER_CPU SYSTEM_CPU DLE_CPU do $ORACLE_HOME/bin/sqlplus -s perfstat/perfstat@iasdb<<EOF insert into perfstat.stats\$vmstat values ( sysdate, $SAMPLE_TIME, '$SERVER_NAME', $RUNQUE, $PAGE_IN, $PAGE_OUT, $USER_CPU, $SYSTEM_CPU, $IDLE_CPU, 0 ); EXIT EOF done done rm /tmp/msg$$

41. Monitoring Servers with vmstat root> vmstat 5 5 kthr memory page faults cpu ----- ----------- ------------------------ ------------ ----------- r b avm fre re pi po fr sr cy in sy cs us sy id wa 7 5 220214 141 0 0 0 42 53 0 1724 12381 2206 19 46 28 7 9 5 220933 195 0 0 1 216 290 0 1952 46118 2712 40 55 0 5 13 5 220646 452 0 14 1 33 54 0 2130 86185 3014 38 59 0 3 6 5 220228 672 0 0 0 0 0 0 1929 25068 2485 25 49 16 10 Assuming an 8 CPU server:  CPU has enqueues when runqueue (r column) > cpu_count  RAM is paging when scan rate (sr) peaks before page-in (pi)

42. Server exception reports Wed Dec 20 page 1 run queue > 2 May indicate an overloaded CPU. When runqueue exceeds the number of CPUs on the server, tasks are waiting for service. SERVER_NAME date hour runq pg_in pg_ot usr sys idl --------------- -------------------- ---- ----- ----- ---- ---- ---- AD-01 00/12/13 17 3 0 0 87 5 8 Whenever Unix performs a page-in, the RAM memory on the server has been exhausted and swap pages are being used. SERVER_NAME date hour runq pg_in pg_ot usr sys idl ----------------- -------------------- ---- ----- ----- ---- ---- ---- AD-01 00/12/13 16 0 5 0 1 1 98 AD-01 00/12/14 09 0 5 0 10 2 88 AD-01 00/12/15 16 0 6 0 0 0 100 AD-01 00/12/19 20 0 29 2 1 2 98 PROD1DB 00/12/13 14 0 3 43 4 4 93 PROD1DB 00/12/19 07 0 2 0 1 3 96 PROD1DB 00/12/19 11 0 3 0 1 3 96

43. Fix for Server Stress  Overloaded CPU – Offload Task to Another Server – Add CPUs – Add Additional Instances/Servers  Overloaded RAM – Add RAM  Cheap $1k/gig – Reallocate RAM from Other Components

44. RAM Disk Solution  Disk I/O remains the biggest bottleneck  100 gig RAM costs $100k  6,000 times faster than disk for Oracle Your app will still run inefficiently, but it runs 6,000 times faster!

45. UNIX server Monitoring rules:  The UNIX vmstat utility provides a wealth of information about the ongoing performance of the Oracle9iAS server.  The vmstat run queue value (r) can indicate a CPU shortage whenever the run queue exceeds the number of CUs on the server.  The vmstat page in values (pi) can indicate a RAM memory shortage.  You can easily define vmstat extension table to hold historical server information and use a UNIX shell script to periodically collect server performance information.  The UNIX server information can be used to generate alert reports and long-term trend reports.

46. Details on Oracle server Monitoring: Oracle9i UNIX Administration Handbook By Don Burleson  In the OW bookstore!

47. OracleAS Case Study

48. Response-time monitoring example with Forms Server:  Database Time  Forms Server Time  Network Time  Client Time

49. Total Response Time: Forms Server Time Form Server Client Database Server Time Database Client Network

50. OracleAS Case Study: Forms Server Tuning  Using OracleAS form server logs, you can often determine the total end-to-end response time. TSE FSERVER_START 0 0 2507559367308400 TSE DBLOGON_START 0 0 2507559367308400 TSE DBLOGON_END 0 0 2507559461832800 Opened file: /u00/app/oracle/prod/forms/F_LOGIN.fmx TSE FSERVER_END -1 0 2507559939113600 TSE FSERVER_START -1 1344 2507560872293600 TSE DB_START 0 0 2507560872801600 TSE DB_END 0 0 2507560967177800

51. Extend iasdb for performance monitoring: create table FormStats ( FORM_ID VARCHAR2(120), EVENT VARCHAR2(120), FSERVER NUMBER, DBASE NUMBER, NWORK NUMBER, CLIENT NUMBER, DATE DATE) ;

52. Load the table from log data: while (((str = in.readLine()) != null) { if (str.startsWith("TSE")) { Add time to appropriate tier; } if (str.startsWith("# ")) { 1 - Extract Event and Form Name; 2 - Load record into database; 3 - Clear times; } }

53. With the table, reporting is easy: Select Number of Form Events with Database Access Time < 4 seconds SELECT COUNT(*) FROM FormStat WHERE (DBASE)/1000) > 4 AND DATE >= SYSDATE-1 AND DATE <= SYSDATE; Select the Form with the greatest time spent in the Forms Server SELECT Form_ID, FSERVER FROM FormStat Where FSERVER = (SELECT MAX(FSERVER) FROM FormStat);

54. Plotting response time data:

55. 98% Threshold Form,Database, Form,Dbase, Network,Client Network Form,Dbase Seconds Events % Events % Events % < 01 63,318 56.68% 91,088 81.54% 109,580 98.09% < 02 79,244 70.93% 99,264 88.85% 110,798 99.18% < 03 88,512 79.23% 103,028 92.22% 111,364 99.68% < 04 93,640 83.82% 104,994 93.98% 111,556 99.86% < 05 96,900 86.74% 106,184 95.05% 111,630 99.92% < 06 99,036 88.65% 106,950 95.73% 111,660 99.95% < 07 100,740 90.18% 107,484 96.21% 111,676 99.96% < 08 101,954 91.26% 107,902 96.59% 111,682 99.97% < 09 103,016 92.21% 108,240 96.89% 111,690 99.98% < 10 103,778 92.89% 108,490 97.11% 111,698 99.98% < 15 106,074 94.95% 109,226 97.77% 111,708 99.99% < 20 107,216 95.97% 109,604 98.11% 111,708 99.99% < 30 108,432 97.06% 110,000 98.46% 111,708 99.99% < 60 109,834 98.32% 110,552 98.96% 111,710 99.99% Total Events: 111,716 111,716 111,716

56. Top offending Forms: Top 10 Forms and Events that use the most Average Form Server Time with a minimum of 10 executions and greater than 2 seconds for execution. 1. Form: d:\prod\forms\F_END_USER_GENERATED_LETTERS.fmx Event: CLICK F_END_USER_GENERATED_LETTERS BUTTONS SAVE_BTN 1 MOUSE Avg Tm: 5.00 Seconds. Number of Executions: 62 2. Form: d:\prod\forms\F_PC_PICK_RETURNS.fmx Event: CLICK F_PC_PICK_RETURNS BUTTONS PROCESS 1 MOUSE Avg Tm: 4.00 Seconds. Number of Executions: 13 Top 10 Forms and Events that use the most Average Database Time with a minimum of 2 executions and greater than 5 seconds for execution. 1. Form: d:\prod\forms\f_pc_case_maint.fmx Event: CLICK F_DIARY DIARY_TAB_ALLOUT DATE_OF_INCIDENT 9 Avg Tm: 472.00 Seconds. Number of Executions: 2

57. Conclusions  Develop a proactive, time-based performance data collection scheme. Real-time OEM and Aggrespy metrics are of little use.  Optimize by adjusting RAM resources parameters.  Once the system is optimized, server monitoring is critical  Server Load Balancing is Critical to properly scale Oracle9iAS

58. Current Books by Don Burleson My web site: dba-oracle.com E-mail me at : don@burleson.cc