Best Practices For Minimizing Oracle EBS R12 Upgrade Downtime - Final

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime
An Oracle White Paper June 2013
Best Practices for Minimizing Oracle EBusiness Suite Release 12 Upgrade Downtime
Executive Overview ........................................................................... 4 Introduction ....................................................................................... 4 Using this document ...................................................................... 4 Preparation and Planning .............................................................. 5 Resolving Performance Issues ...................................................... 5 Prediction of Performance Issues ...................................................... 6 Summary ........................................................................................... 6 Preparation and Planning - Pre-Upgrade Environment and Activities Preparation and Planning Release 12 Upgrade Activities ........... 7 Resolving Performance Issues - Using Diagnostics ..................... 13 Resolving Performance Issues Common Solutions .................. 16 Key Resources and Documents .................................................. 19 Preparation and Planning - Pre-Upgrade Environment / Activities ... 22 OATM (Oracle Applications Tablespace Model) .......................... 22 Multiple Organizations (Multi-Org) Architecture ........................... 22 Purge old data ............................................................................. 22 Purge interface data .................................................................... 23 Gather CBO Statistics ................................................................. 23 Fixed Object and Dictionary Statistics.......................................... 24 Drop MRC Schema ..................................................................... 24 Preparation and Planning Release 12 Upgrade Activities ............. 24 Performance Fixes ...................................................................... 24 Removing Unnecessary Workloads / Overheads ......................... 25 Upgrade and Initialization Parameters ......................................... 29 Dynamic / Hyper Threading ......................................................... 36 Gathering CBO Statistics ............................................................. 37 Resolving Performance Issues ........................................................ 46 Diagnostics to be gathered during the upgrade ........................... 46 Useful Scripts .............................................................................. 59
Common Solutions ...................................................................... 68
Executive Overview
This document offers a strategy for upgrading Oracle E-Business Suite to Release 12 that minimizes downtime, whilst minimizing the implementation and resource required. The content is specifically aimed at upgrades from 11.5.10CU2 to R12.1.3, but it will also be relevant for upgrades from 11.5.10CU2 to R12.2. Some of it will be relevant for upgrades between intermediate versions. In particular it should allow you to: Minimize downtime during the final iteration. Minimize number of test iterations used to resolve performance issues. Identify issues early. Ideally on the first test iteration. Simplify the final iterations, so that they require the minimum of resource
Introduction
Using this document
The summary section contains the high level recommendations and should be reviewed fully. The detailed Preparation and Planning and Resolving Performance Issues sections contain detailed explanations for the recommendations or detailed instructions. They may also contain important notes that you should be aware of. This section will typically be used for reference.
There are two parts to the suggested approach: Preparation and Planning Resolving Performance Issues
Both are these are discussed below.
Preparation and Planning

This includes: Recommended configuration or setup. Tasks that should be run as part of the Release 12 upgrade process. Diagnostics that should be captured for all runs. Advice on monitoring general performance. Advice on determining upgrade parameters (AD Parallel workers, batch size, max_parallel_servers) for each implementation. This will typically apply to all implementations.
Resolving Performance Issues

This will include advice on how to resolve specific performance issues that are encountered on each implementation.
Our experience is that although there are some specific areas and jobs which sometimes have performance issues, there are often new undiscovered / unreported performance issues. It is not possible to predict where performance issues will occur for a given customer.
Prediction of Performance Issues

It can be difficult to predict how long a Release 12 upgrade will take (in terms of downtime). In particular, it is difficult to produce reliable or accurate estimates based on database size. The downtime is dependent on the modules and functionality used, and on the hardware specification. The nature of the cost based optimizer and the mathematics behind contention (e.g. queuing theory) means that jobs that previously had no known performance issues can suddenly have performance issues for particular customers. For example: Very small changes in CBO statistics, database initialization parameters and other configuration could lead to a different execution plan being used. Queue (wait) times can start to increase rapidly (exponentially) once the load reaches a certain level.
The modules, data distributions, and functionality used, not only across different tables and entities, but chronologically too, can all mean that the time to run a particular upgrade job or SQL varies enormously; or a different execution plan is used. The execution plan that was efficient for one customer may not be for another customer. The parameters or configuration of the environment may cause contention when previously they didnt: in this context, small changes can have a large impact. Similarly, a performance issue encountered on a job may not be the same as a previous issue on the same job. On the other hand, some jobs may run long for all customers that use certain modules heavily.
Summary
This section contains a summary of the recommendations to minimize downtime. More details, explanations and instructions can be found in the detailed sections that follow.
Preparation and Planning - Pre-Upgrade Environment and Activities
This includes steps to prevent performance issues and reduce downtime. These are all activities to be carried out on the 11.5.10 environment prior to Release 12 upgrade.
List of recommended pre-upgrade activities
Convert to OATM (Oracle Applications Tablespace Model). Convert to the new Multiple Organizations (Multi-Org) architecture. Purge all old data that is no longer needed prior to the upgrade. Flush all the interfaces, such as Auto Invoice, Journal Entry Import, Order Import etc. Drop MRC Schema if it still exists. Gather the schema statistics (with GATHER_AUTO option for all schemas) close to the start of the downtime. Use FND_STATS or Gather Statistics concurrent programs. Gather Fixed Object and Dictionary Statistics
Preparation and Planning Release 12 Upgrade Activities

Performance Fixes
Ensure that any recommended performance patches are applied. Check the following: Oracle E-Business Suite Pre-install Patches Report [Video] (Doc ID 1448102.1) Oracle E-Business Suite Recommended Performance Patches (Doc ID 244040.1)
Also ensure that you pre-install patch 16410424.R12.FND.B for the R12.1.1 upgrade. This contains the latest version of FND_STATS.
Removing Unnecessary Workloads / Overheads
Disable any custom triggers and business events Disable auditing if enabled. Review and disable all debug or logging; do this at all levels (site, responsibility, user level etc.). If possible run in noarchivelog mode. Disable flashback DB.
Remove TDE (Transparent Data Encryption) from high volume tables. Consider running AutoConfig in parallel on a multi-node system. Split any RDBMS Upgrade, Platform Upgrade, Conversion to OATM into a separate downtime period. Use TUMS (The Upgrade Manual Script) to avoid running tasks not relevant to your system Minimize Historical Data To Be Upgraded (Upgrade by Request) Use Staged Application System (APPL_TOP) Parallelize pre and post upgrade technical activities Define separate concurrent manager queue for post upgrade jobs Using Distributed AD and Shared APPL_TOP may give benefits for smaller DBs.
Note that Defer Upload of Patch Information (See My Oracle Support Knowledge Document 225165.1) cannot normally be used, as subsequent patches may rely on information from previous patches, and the Release 12 upgrade is a series of patches (e.g. R12.1.3 RUP follows R12.1.1 and uses patch information from R12.1.1).
Upgrade and Initialization Parameters
For 32 cores or fewer initially set: parallel_max_servers = 2 x number of CPU cores. AD Parallel workers start with 1 x number of CPU cores. Possibly increase to 1.5 x number of CPU cores. job_queue_processes = number of CPU cores parallel_max_servers = 1 x number of CPU cores. AD Parallel workers = between 0.5 and 1.0 x number of CPU cores.
For more than 32 cores, start with:
Based on the performance diagnostics, you may need to decrease these values or you may be able to increase them. The level of contention and resource (CPU) usage (in AWR) will be the key to determining this.
Be careful if the hardware has dynamic or hyper-threading. In such cases the number of logical CPUs could be a multiple of the number of cores. If so ensure you calculate the values above using the number of CPU cores and not the number of logical CPUs. You should only increase the number of AD Parallel workers if the level of contention on all long running AD Parallel jobs is low. Similarly, you should only increase parallel_max_servers if the level of contention on all parallel query/DML jobs is low. Typically, this means that you should resolve SQL Tuning (poor execution plans) and Contention issues before increasing the AD Parallel workers or parallel_max_servers. Set AD Parallel batch size to 10,000. Note: Small extents (e.g. 128k = 16 blocks) will seriously limit the batch size. You are likely to end up with batches of much less than 1000, regardless of what batch size you select. If possible you should maximize SGA and PGA sizing. Check feedback from AWR. Some starting rules of thumb are: log buffer = 30 to 100 Mb shared pool = 1 to 4 GB pga target = 3 to 20 GB SGA/buffer cache = multi GB: be generous without causing excessive paging If specified, remove db_file_multiblock_read_count. This is the recommended value for normal running of Oracle E-Business Suite. optimizer_dynamic_sampling level should normally be 2 (which is the default if not set) or higher. We recommend a value of 4 during the Release 12 Upgrade, but you must revert to 2 (or remove) after the upgrade. Note that the values of the initialization parameters above (except db_file_multiblock_read_count) may be different from the values used for normal running. So be sure to revert after the Release 12 upgrade has completed. For other initialization parameters, refer to My Oracle Support Knowledge Document 396009.1, Database Initialization Parameters for Oracle E-Business Suite Release 12.
Resolve any SQL Tuning Issues on Long Running Jobs Early
SQLs with poor execution plans often exhibit disproportionate use of a particular resource (e.g. I/O, CPU or memory) or a high level of contention (e.g. on blocks in buffer cache or on I/O). This can be misleading if you are trying to resolve general performance issues or investigating the benefit of changes in architecture or configuration. Sometimes these architectural or configuration changes can cause regressions on long running jobs with poor execution plans and high resource usage or contention. And these regressions outweigh the benefits elsewhere. Once you resolve any SQL tuning issues on long running jobs, you are much better placed to investigate and assess the benefits of changes in architecture, configuration and numbers of AD Parallel Workers/Parallel Max Servers.
Dynamic/Hyper Threading
Be aware of Dynamic or Hyper Threading. We recommend configuring dynamic/hyper-threading for maximum throughput per CPU cycle i.e. 1 thread for each core. On Oracle SPARC T4-4 you can do this by setting dynamic threading to max-ipc. In addition, you will need to disable any power management. Power management will either prevent threading for maximum throughput, or throttle the CPU capacity.
Gathering CBO Statistics
Gathering Schema Statistics
In normal circumstances you should not need to manually gather schema statistics during the Release 12 upgrade. The script adsstats.sql automatically gathers them towards the end of the R121.1 upgrade (e.g. R12.1.1 phase: last+63). However, you should gather CBO statistics for all Oracle E-Business Suite schemas with GATHER_AUTO option using FND_STATS (or gather statistics concurrent program) again, after the entire Release 12 upgrade, but prior to the system being online and available to users. If the adsstats.sql script is taking a significant amount of time, you could reduce the upgrade time by: Exporting schema statistics gathered during test runs (by adsstats.sql - at same point: phase: last+63). Importing these statistics instead of running adsstats.sql.
10
Only do this if you estimate that the time saved will be significant, and you have checked that parallel execution is being used effectively (with parallel_max_servers set to a suitable value, such as 2 x number of cores). Use the recommended scripts in the detailed section. If the adsstats.sql job is taking a long time during the R12.1.1 upgrade there are two further approaches that may help for large environments, where there are tables with more than 100 million rows. Use dbms_stats.auto_sample_size in adsstats.sql Sample specific long running tables at a lower percentage
See the detailed section for more information. Once you go live, watch out for statistics that adsstats.sql has unlocked and which may need to be locked again. Note that adsstats.sql is often confused with adstats.sql, which disables automatic statistics gathering and gathers fixed object and dictionary statistics. Ensure that you pre-install patch 16410424.R12.FND.B for the R12.1.1 upgrade. This contains the latest version of FND_STATS.
Fixed Object and Dictionary Statistics
Fixed Object Statistics should be gathered:
After any associated platform or database upgrade that is part of the overall Oracle EBusiness Suite upgrade. After any SGA/PGA parameters have changed. After Release 12 upgrade, when there is representative activity on the system.
Dictionary Statistics should be gathered:
After any associated platform or DB upgrade that is part of the overall Oracle E-Business Suite upgrade. After the Release 12 upgrade. After move to OATM.
They are not gathered automatically in the Release 12 upgrade by adstats.sql or any other method.
11
There may be additional specific circumstances during the upgrade where you need to gather fixed object or dictionary statistics (such as before importing schema statistics or running SQLT or AWR reports when AWR has grown significantly). See the detailed section for more information. Do not use adstats.sql to gather fixed object and dictionary statistics (unless as part of a database upgrade). Use the APIs directly, as instructed in detailed section below.
Diagnostics to be gathered during the upgrade
Automatic Workload Repository (AWR) should be enabled with a default snapshot of 30 minutes. For short upgrades, a shorter snapshot may be more suitable. The AWR retention period should be long enough to cover the duration of the upgrade run and a significant period afterwards (to gather diagnostics and analyze). The suggestion is N+7 days, where N is the estimated upgrade time, but a longer period will provide more time to gather subsequent diagnostics and statistics. We very strongly advise that the statistics level is set to ALL for the duration of the Release 12 upgrade test runs. When analyzing Release 12 Upgrade performance issues, your goal is to: Prevent wasted test iterations. You should aim to provide solutions that solve the issue first time. Maximize the number of performance issues resolved.
To do this we recommend using the steps outlined in the Express Diagnosis of Oracle E-Business Suite Release 12 Upgrade Performance Issues whitepaper. The key steps are: Before the Oracle E-Business Suite Upgrade: Set the statistics level to ALL (there are no actual statistics without this) During the Oracle E-Business Suite Upgrade: Identify top SQL in AWR (see useful scripts) Obtain Display Cursor Reports (ALL +ALLSTATS) for the long running SQLs Obtain SQL Monitor Reports for SQL that uses Parallel Query or DML Identify when a piece of SQL ran (see useful scripts)
12
Match long running SQL with a job (see useful scripts) Report on CBO Statistics for all Oracle E-Business Suite tables (see useful scripts) Obtain AD Job Timing Reports Identify long running upgrade jobs (See useful scripts) Obtain the file versions for long running jobs Obtain AWR Reports Run afxplain.sql to obtain detailed CBO statistics, database parameters, CBO Parameters and metadata; do this for individual pieces of SQL
After the Oracle E-Business Suite Upgrade
For more detailed analysis, you will require the following: AD utility and worker logs SQLT with XTRACT for long-running SQL
In addition, Oracle Support and Development may require: AWR Export AD Parallel tables export
See the Diagnostics section in Resolving Performance Issues below for more details.
Resolving Performance Issues - Using Diagnostics

File Versions
To find out versions of files actually used in an upgrade then check the unified driver. Do not rely on the version of the file or object in the file system or database, as it could be a later version applied by a subsequent patch.
Statistics Level = ALL
This is the simplest way to see actual execution statistics (including elapsed time, physical reads, buffer gets) for each execution plan line (on SQLT and Display Cursor report). The alternative of SQL Trace and TKPROF requires editing standard code.
13
This should allow identification of the root causes of issues on the first iteration, and will speed up resolution. Note that setting statistics_level to ALL while AWR is enabled could significantly increase the use of the SYSAUX tablespace (table WRH$_LATCH_CHILDREN). So monitor the usage of this tablespace.
AWR Export
The AWR Export allows the AWR tables to be loaded into another environment. They can then be analyzed to discover where waits occur (job, SQL, object etc) and the patterns.
AWR Reports
You should obtain AWR reports for the period that the upgrade is running. You should also obtain AWR reports for the duration of long running jobs. These will allow you to identify: Long running SQL (SQL Statistics/SQL Ordered By) Contention and bottlenecks (Top 5 Timed Foreground Events/Foreground Wait Events) CPU utilization
If you see high levels of a particular wait(s), you should first check to see if it only occurs with particular long running SQL and jobs, before you assume it is a system wide configuration or resource issue. The Active Session History report for particular SQL can also be useful in identifying the waits/events for each row source or object. Note that you should gather fixed object and dictionary statistics before running AWR reports, especially if you have set statistics level to ALL or you have changed to a high retention period or a short snapshot interval.
AD Parallel tables
The AD_PARALLEL_UPDATES, AD_PARALLEL_UPDATE_UNITS tables can give information on the actual rows processed, the number of batches, progress over time, and long running batches (that might indicate locks/sleeps or data distribution issues).
14
AD_TASK_TIMING gives start and end times of jobs and workers, which can help identify all long running jobs, and match long running SQL and performance issues (on AWR) with specific jobs. It also helps identify low worker utilization (and potentially resource utilization) due to phasing waits.
AD Job Timing Reports
The job timing report (adtimrpt.sql) reports the top 100 time consuming jobs. It also reports the timing of each upgrade phase and failed, deferred, re-started and skipped jobs. The detailed report adtimdet.sql (adt.lst) reports on all jobs by phase and by elapsed time. Note that the Top 100 Time Consuming Jobs section of the standard adtimrpt.sql report lists all workers for AD Parallel jobs separately. So the top 100 can be dominated by a handful of jobs. An alternative is to use the SQL in the Long Running Upgrade Jobs section of Useful Scripts.
AD Utility and Worker Logs
The AD utility and worker logs can also be useful for diagnostics, giving you more detail about what happened and when. The AD workers logs (adworknnn.log) will give you the activities carried out by each worker and the timings.
SQL Trace Job Specific
If you are unable to obtain the diagnostics you need to resolve a performance issue from the display cursor, SQL monitor, and AWR, you should obtain a 10046 SQL Trace (with waits level 8) for the job on the next test run. It is advisable to enable SQL Trace for the specific job (by temporarily editing the SQL script). See detailed section for instructions.
SQL Specific
Once you have identified the long running SQL (in the long running jobs), the following will provide diagnostics for individual pieces of SQL: Display Cursor Report SQL Monitor Report (for SQLs that use Oracle Parallel Query/DML afxplain.sql (provides detailed CBO Statistics, DB/CBO Parameters and Metadata for a SQL) SQLT Output using the XTRACT method.
15
The first three can be provided during the upgrade run, and the last after it has completed.
Long Running SQL, Contention and Tuning
For long running jobs or SQL it is best to start by investigating if good execution plans are being used. A poor execution plan (or even just one that is moderately sub-optimal) can be the root cause of contention, especially if that contention only occurs during a particular job. Once you have removed any unnecessary contention caused by sub-optimal execution plans, a small amount of contention (e.g. 5 to -15% on particular waits) between AD Parallel or Parallel Execution sessions can be a useful indicator that you are getting the most out of the available resources. Be aware that although the entire wait time of some wait events can be classed as contention, this is not the case for all waits, and in particular not for I/O waits such as db file sequential read/db file scattered read and direct path reads/writes.
Resolving Performance Issues Common Solutions

Known Issues
Once you have identified the long running jobs and SQL, you can check My Oracle Support for known issues and potential solutions or workarounds. However, bear in mind that the fix or workaround may not necessarily fix your particular problem. If you cannot confirm that you have exactly the same issue (from the diagnostics) you may still apply the fix, but should continue to gather diagnostics and search for a solution until the issue is fully resolved.
Custom Indexes
Create custom indexes for long running jobs where a new index could significantly improve the execution plan and performance.
SQL Profiles for Inefficient Execution Plans
If you have identified that a long running job has an inefficient execution plan, you could use an SQL Profile to apply hints that will help the CBO choose a better execution plan. You will need SQL tuning expertise to do this.
16
Pragmatic Stable Execution Plans
A pragmatic stable execution plan may not be the very best execution plan possible, but it will not be inefficient and the job time will be predictable and stable. In most cases, there is one join order that will give a good execution plan and minimize throwaway. For AD Parallel jobs, a pragmatic execution plan will lead with the driving tables and use nested loop joins and index access.
Long Running Index Creation
This will typically apply to xdf and odf jobs. You should obtain an AWR report for the snapshots where the index is being created, then investigate further. Although pre-creation can help in a few circumstances, there is more likely to be another solution. If you do pre-create, take care to do so correctly. See detailed section.
High level of contention on indexes on long running DML jobs
If a long running job runs heavy DML, has a high level of contention and that contention is on particular indexes then you should consider dropping the indexes before the job and recreating them afterwards (provided the index is not used in the meantime). You should ensure that you recreate in parallel and with exactly the same definition. And remember to ALTER INDEX to revert the degree of parallel (NOPARALLEL).
High Level of enq: HW contention or enq: HV contention
If a long running job inserting into a table and indexes has a high level of waits enq: HW contention or enq: HV contention, you could do the following: If the wait is occurring largely on particular indexes, drop the indexes before the job and recreate them afterwards (provided the index is not used in the meantime). Increase the extent size on the table and indexes. Pre-allocate extents for the table and indexes.
17
Partition the table and any indexes to spread the load across multiple partitions. Note that the tables and indexes would still be partitioned after go live. So only do this if it will also give definite benefits after going live on the production environment.
High Level of redo log waits log buffer space, log file sync etc.
If you have a high level of waits associated with redo log, especially log buffer space and log file sync, you could consider:
changing the configuration of redo logs move to faster filer increase the size; increase the number or increase the log parallelism (hidden parameter _log_parallelism_max).You could consider running with NOLOGGING, but this is not advised.
Long Running Statistics Gathering (adsstats.sql)
If the upgrade is taking a long time to gather CBO statistics (adsstats.sql), then you can consider the following strategies: Increasing parallel_max_servers and pga_aggregate_target. (adsstats.sql should run with few or no concurrent tasks). Using all nodes on an Oracle RAC system. Importing statistics gathered during test runs (covered in Preparation and Planning section).
Gathering or Deleting Statistics to resolve specific performance issues
Performance issues may indicate the need for you to gather or delete statistics for specific objects. The specifics of doing so will depend on the exact circumstances, and the solutions may not be simple. See the detailed section.
Long Running Materialized View xdf/odfs
If you have long running xdf or odf jobs creating materialized views (MV), consider cleaning up or truncating any large MV logs (note that this requires MV complete refresh).
Long Running Jobs that might not be needed
18
Check if any long running jobs are actually needed on your system (e.g. for a module or localization that you do not use or are not licensed for). Oracle Support may advise that you can skip a job, or offer a fix or workaround for your case.
Maximum AD Parallel Workers
The Auto Patch utility specifies a maximum number of workers. The number depends on several factors, and under certain circumstances may be less than the number of workers you require. See the detailed section for how this maximum is calculated, and how you might resolve the issue (increase processes parameter, or reduce other sessions/processes).
Key Resources and Documents

Oracle E-Business Suite Upgrade Guide Release 11i to 12.1.1 (E16342-04 Aug 2010) Especially Planning for an Upgrade (Chapter 1) and appendices Reducing Downtime (E) and Upgrade By request (G) Planning Your Oracle E-Business Suite (EBS) Upgrade from Release 11i to Release 12 (Doc ID 1406960.1) Database Preparation Guidelines for an E-Business Suite Release 12.1.1 Upgrade (Doc ID 761570.1) EBPERF FAQ - Collecting Statistics with Oracle EBS 11i and R12 (Doc ID 368252.1) Using a Staged Applications System (APPL_TOP) to Reduce Patching Downtime in Oracle EBusiness Suite Release 12 (Doc ID 734025.1) Sharing The Application Tier File System in Oracle E-Business Suite Release 12 (Doc ID 384248.1) AD Command Line Options for Release R12 (Doc ID 1078973.1) Oracle E-Business Suite Recommended Performance Patches (Doc ID 244040.1) Oracle Applications Release 12 Upgrade Sizing and Best Practices (Doc ID 399362.1) Oracle E-Business Suite Pre-install Patches Report [Video] (Doc ID 1448102.1) Oracle E-Business Suite Upgrades and Platform Migration (Doc ID 1377213.1) EBS Database Upgrade to 11g Performance Best Practices (Doc ID 1100303.1)
19
Oracle Applications Tablespace Model Release 11i - Tablespace Migration Utility (Doc ID 248857.1) Use of Multiple Organizations In Oracle Applications Release 11i (Doc ID 210193.1) How to Send a File to Oracle Support Using FTP (ftp.oracle.com) (Doc ID 464666.1) SQLT (SQLTXPLAIN) - Tool that helps to diagnose a SQL statement performing poorly or one that produces wrong results (Doc ID 215187.1) How Does Adpatch Determine The Number Of Workers To Recommend? (Doc ID 800024.1) Oracle Database VLDB and Partitioning Guide 11g Release 2 (11.2). Part Number E25523-01, How Parallel Execution Works and Tuning General Parameters for Parallel Execution sections. Using AutoConfig to Manage System Configurations in Oracle E-Business Suite Release 12 (Doc ID 387859.1) section 5.1. Running AutoConfig in Parallel Across Multiple Nodes. How to verify or create a Database Object using an odf (adodfcmp) or xdf (FndXdfCmp) file? (Doc ID 551325.1) The ADODFCMP Utility (Doc ID 137176.1)Patching Best Practices and Reducing Downtime, Note 225165.1 Database Initialization Parameters for Oracle E-Business Suite Release 12 [ID 396009.1] Best Practices for Gathering Optimizer Statistics. This is available on the Oracle Optimizer blog: https://blogs.oracle.com/optimizer/ and Oracle Technet: http://www.oracle.com/technetwork/database/bi-datawarehousing/dbbi-tech-info-optmztn092214.html Understanding Optimizer Statistics. This is available on Oracle Technet: http://www.oracle.com/technetwork/database/focus-areas/bi-datawarehousing/twpoptimizer-stats-concepts-110711-1354477.pdf Upgrading from Oracle Database 10g to 11g: What to expect from the Optimizer. This is available on Oracle Technet: http://www.oracle.com/technetwork/database/focus-areas/bidatawarehousing/twp-upgrading-10g-to-11g-what-to-ex-133707.pdf Fixed Objects Statistics Considerations (Doc ID 798257.1) Managing CBO Stats during an upgrade to Oracle 10g or Oracle 11g (Doc ID 465787.1)
20
bde_last_analyzed.sql - Verifies CBO Statistics (Doc ID 163208.1)
21
Preparation and Planning - Pre-Upgrade Environment / Activities

This includes steps to prevent performance issues and reduce downtime. These are all activities to be carried out on the 11.5.10 environment prior to Release 12 upgrade.
OATM (Oracle Applications Tablespace Model)

Chapter 1 (Planning for an Upgrade) of the Oracle E-Business Suite Upgrade Guide describes how you can convert to the new Oracle Applications Tablespace Model (OATM) while you are still on Oracle E-Business Suite 11i. See Oracle Applications Tablespace Model Release 11i - Tablespace Migration Utility (Doc ID 248857.1) and Planning Your Oracle E-Business Suite (EBS) Upgrade from Release 11i to Release 12 (Doc ID 1406960.1)
Multiple Organizations (Multi-Org) Architecture

Chapter 1 (Planning for an Upgrade) of the Oracle E-Business Suite Upgrade Guide, describes how you can convert to the new Multiple Organizations (Multi-Org) architecture while you are still on Oracle E-Business Suite 11i. See Use of Multiple Organizations In Oracle Applications Release 11i (Doc ID 210193.1) and Planning Your Oracle E-Business Suite (EBS) Upgrade from Release 11i to Release 12 (Doc ID 1406960.1)
Purge old data

Purge all old data that is no longer needed prior to the upgrade. This can be quite a lengthy process, and you may need to resolve issues around performance, execution frequency and the periods to be purged in each execution. So this should be started as soon as possible. There are over 100 standard purge programs in 11i. You can use OAM to administrate, configure, initiate and monitor purge programs. System Administrator >Oracle Applications Manager >Purging/Critical Activities.
22
A list of purge programs is listed in document Reducing Your Oracle E-Business Suite Data Footprint using Archiving, Purging, and Information Lifecycle Management (Doc ID 752322.1). This contains Archive_Purge_ILM_paper[1].pdf
Purge interface data

Flush all the interfaces, such as Auto invoice, Journal entry import, and order import.
Gather CBO Statistics

Regularly gather schema statistics with GATHER_AUTO option for all schemas while still on Release 11i. Then gather schema statistics (with GATHER_AUTO option for all schemas) close to the start of the Release 12 upgrade. Always use the FND_STATS package or the Gather Statistics concurrent programs. Do not use the DBMS_STATS package or Analyze.
Zero or Incorrect Statistics on transient/temporary tables
Be aware that the statistics from the 11.5.10 environment may contain some issues, even if they have been gathered with the Auto option. You should check the statistics for transient, temporary or Global Temporary (GT) tables. These are tables which can alternately contain rows for processing (e.g. import/interface/staging), or contain no (or very few) rows. For GT tables, rows are only visible to the session that inserted them, and those rows only persist for the duration of the session. If you gather statistics on these tables, you will only collect statistics for the rows present in that table for your session: that is likely to result in statistics for 0 rows. In addition, the statistics for GT tables are not session specific, so there is only one set of statistics and these are visible to all other sessions. Even if the statistics are gathered by a session when the tables are populated, they could still result in suboptimal execution plans if the volume or distribution of data is different for other sessions. Similarly, for other transient, temporary or staging tables, you may have gathered statistics when the table is empty.
23
Having zero statistics (or statistics with low row count) could result in poor execution plans for temporary/transient tables that contain a large number of rows. Ideally, all these temporary, transient or global temporary tables should have no statistics, and dynamic sampling should be used. However, in specific cases, a strategy of populating with indicative statistics (when the table is populated) or gathering statistics each time a process is run may have been taken. In some cases the seeded Oracle E-Business Suite code will gather statistics.

These should have been gathered and be correct/up to date on the pre-upgrade environment.
Drop MRC Schema

All programs and reports now use the APPS schema. The MRC_APPS schema is no longer needed, so dropping it frees space and reduces processing overhead during the upgrade. You can drop this schema prior to the Release 12 upgrade if it still exists. It should have already been removed on upgrade to 11.5.10. See Oracle E-Business Suite Upgrade Guide.
Preparation and Planning Release 12 Upgrade Activities

Note: Only run the Release 12 upgrade on one Oracle RAC node. The majority of the elapsed time of the Release 12 upgrade is taken by jobs running DML (INSERT, UPDATE, DELETE). These jobs use multiple workers/parallel severs, which all access the same objects and blocks at the same time. So the additional communication between nodes on the cluster (and cluster waits) will significantly outweigh the gains from using the additional CPUs to increase throughput.
Performance Fixes
Review Oracle E-Business Suite Pre-install Patches Report [Video] (Doc ID 1448102.1) Oracle E-Business Suite Recommended Performance Patches (Doc ID 244040.1)
24
Note that 244040.1 only contains Database, Applications Technology and Post Release 12 go live product patches. It does not include fixes for product performance issues during the Release 12 upgrade: these are available as part of consolidated upgrade patches or pre-upgrade patches (see Note 1448102.1). Fixes of particular note are: Forms/reports generation regression with 11g: 8557019 Optimizer:Dynamic Sampling on indexes 12942119 (11.2.0.2) AD-Parallel improved scalability: 8917381 (Part of 9179588 AD CUP)
Other recommended performance fixes (if not already applied) are:
Also ensure that you pre-install patch 16410424.R12.FND.B for the R12.1.1 upgrade. This contains the latest version of FND_STATS.
Removing Unnecessary Workloads / Overheads

To reduce the workload you should do the following: Disable any custom triggers and business events. Disable auditing if enabled. The Oracle E-Business Suite Upgrade Guide states that the Oracle AOL Audit Trail should be disabled before upgrade, anyway. Review and disable all debug and logging that has been enabled using profiles. Do this at all levels (e.g. site, responsibility, user level.). If possible, run in noarchivelog mode. Disable flashback DB. Remove TDE (Transparent Data Encryption) from high volume tables. Consider running AutoConfig in parallel on a multi-node system. See note Using AutoConfig to Manage System Configurations in Oracle E-Business Suite Release 12 (Doc ID 387859.1) section 5.1. Running AutoConfig in Parallel Across Multiple Nodes.
25
Note that Defer Upload of Patch Information (See note 225165.1) cannot be used as subsequent patches may rely on information from previous patches and the Release 12 upgrade is a series of patches (e.g. R12.1.3 RUP follows R12.1.1 and uses patch information from R12.1.1).
Split Into Separate Steps
Identify tasks that could be completed in a separate downtime period, prior to the production upgrade. The following could all be candidates: Upgrade Database version (to latest certified for the current Oracle E-Business Suite level) Convert to Oracle Applications Tablespace Model (OATM). Other planned platform (hardware or OS) upgrades. Tasks identified in the Oracle E-Business Suite Upgrade Guide Planning for an Upgrade (Chapter 1) and appendices Reducing Downtime (E) and Upgrade by Request(G). The latter includes the topic covered in the section below and will include pre-upgrade and post-upgrade tasks.
Use TUMS (The Upgrade Manual Script) to avoid running tasks not relevant to your system
The TUMS report lists tasks that you can omit from the upgrade because they do not apply to your system (for example, a task required for a product that you do not use or a patch that you have previously applied). TUMS is delivered in a patch (7705743), which supplies the scripts you need to examine your system and create the report. We strongly recommend you create and review the TUMS report before you begin the upgrade. Download and apply patch 7705743, then run the script adtums.sql to generate the report tumsr12.html. The tumsr12.htm report lists the steps (identified by the TUMS step key in this book) that do not apply to your installation. You may safely ignore any steps listed in this report. See Oracle E-Business Suite Upgrade Guide Preparing for the Upgrade (Chapter 2).
Minimize Historical Data To Be Upgraded
See Oracle E-Business Suite Upgrade Guide appendix Upgrade by Request (G). Also see R12.0 and R12.1: FAQ for the SLA Upgrade: SLA Pre-Upgrade, Post-Upgrade, and Hot Patch (Doc ID 604893.1)
26
Note that there are some SLA upgrade jobs in Release 12, especially for Oracle Payables, which still process all the historical data, even if you have only chosen to upgrade the most current data. The Release 12 upgrade migrates all the transactions for the following regardless - Payables (Transaction Entities, Events, Journal Headers, Journal Lines), Receivables (Transaction Entities). See note Oracle Applications Release 12 Upgrade Sizing and Best Practices (Doc ID 399362.1) section "Post Upgrade - Historical Financial Documents (optional)"
Staged Application System (APPL_TOP)
A Staged Applications system represents an exact copy of your Production system, including all APPL_TOPs and a copy of the Production database. Patches are applied to this Staged system, while your Production system remains operational. When all patches have been successfully applied to the Staged system, the reduced downtime for the Production system can begin. The Staged APPL_TOP is used to run the database update into the Production database, as well as synchronizing the Production APPL_TOP. See My Oracle Support note Using a Staged Applications System (APPL_TOP) to Reduce Patching Downtime in Oracle E-Business Suite Release 12 (Doc ID 734025.1) Note that for the upgrade from 11.5.10CU2 to R12.1.3 this approach can only be used to upgrade directly from Release 11.5.10CU2 to R12.1.1. You must complete the upgrade to 12.1.1 completely, including the post install steps, before you apply the 12.1.3 Rollup. Theoretically you can upgrade from R12.1.1 to R12.1.3 using a Staged Applications System; however it is unlikely to save any time, as you need to complete the upgrade to R12.1.1 first (including DB upgrade and synchronizing patch histories). For the Oracle internal GSI Release 12 Upgrade, Oracle created a Master Release 12 APPL_TOP prior to the upgrade downtime using our most recent test upgrade, with all file system patches preapplied. In addition, generation of files such as forms, reports and message files was performed before the upgrade downtime.
Using Distributed AD with Shared APPL_TOP
This may give some benefits for customers with smaller DBs. See notes 384248.1 and 236469.1 There will be no need to apply the same patch on multiple tiers.
27
Distributed AD adds to the degree of parallelism by distributing AD workers across application tier nodes and improves timing for the parts of the driver that have a lower DB load (i.e. applying code, metadata etc.) However, to obtain these benefits you need to increase the number of workers and (if you are already at the optimum number of workers) this will significantly increase contention on those jobs with heavy DB load.
Parallelize pre and post upgrade technical activities
Where practical, perform technical upgrade tasks in parallel. For example, while the upgrade driver is running on the database, you could perform Release 12 setups on the application tiers, or re-register single sign-on.
Define separate concurrent manager queue for post-upgrade job
There are more than 50 distinct concurrent programs automatically submitted in the downtime portion of the upgrade to Release 12. Many of these programs will run in multiple threads, so the total number of concurrent requests that form part of the post-upgrade step is much higher. These programs will be picked up and executed by the concurrent manager once the system is up. Consequently, their execution will be mixed with the execution of ongoing concurrent jobs in the system. You may find it beneficial to define a separate concurrent manager queue for these requests, to enable post-upgrade testing of standard run-time requests to proceed without interference from upgraderelated processes. Such a definition can be achieved by using inclusion and exclusion rules to prevent other manager queues such as standard from picking up these requests, and allow this new manager queue to process these requests only. Doing this now allows you to control the number of target processes allocated to these post-upgrade concurrent programs, including doing this dynamically with the use of work shifts. For additional details on configuring new manager queues, target processes, inclusion/exclusion rules and work shifts, refer to the Applications System Administrator's Guide from Oracle E-Business Documentation Library. See Oracle Applications Release 12 Upgrade Sizing and Best Practices (Doc ID 399362.1) Best Practices section.
28
Upgrade and Initialization Parameters

Note that the values of the initialization parameters below (except db_file_multiblock_read_count) may be different from the values used for normal running. So be sure to revert after the Release 12 upgrade is complete. For other initialization parameters, refer to note 396009.1 Database Initialization Parameters for Oracle E-Business Suite Release 12.
AD Parallel Workers and parallel_max_servers
The number of parallel execution threads is determined by the initialization parameter parallel_max_servers. This will impact the elapsed time of jobs that use parallel SQL and DML. Such jobs include apintbal.sql, apxlaupg.sql, and adsstats.sql. The number of workers used in the AD Parallel Utility is also configurable. It is important to get the number of AD Parallel workers and the parallel_max_servers right. Our typical recommendation is to set: parallel_max_servers = 2 x number of CPU cores. AD Parallel workers start with 1 x number of CPU cores. Possibly increase to 1.5 x number of CPU cores. The above values are recommended to maximize CPU utilization. However, the constraints may lie in I/O, memory, or elsewhere. So check memory utilization (no swapping/excessive paging) and I/O response times (histogram shows a fat tail for the longer wait times). Also check for buffer busy waits spread across all objects (high levels on specific objects can often indicate a sub-optimal execution plan). The number of CPUs/cores on systems is increasing, and consequently the number of CPUs presented to the operating system is also increasing. This is particularly true if the hardware has dynamic/hyperthreading capabilities and you do not disable them for the Release 12 upgrade.
However, note that lower values may turn out to be more optimal, for the following reasons:
The recommendations above are usually valid if the number of CPU cores is fewer than 32. However, once you get to 32 cores and beyond, the levels of contention can increase significantly, outweighing any gains in CPU and resource utilization obtained from increasing the workers/threads.
29
So, if the number of cores is 32 or more, then you may need to start with parallel_max_servers and AD Parallel workers below the above levels. For example: parallel_max_servers = 1 x number of CPU cores. AD Parallel workers to between 0.5 and 1.0 x number of CPU cores.
Be careful if the hardware has dynamic or hyperthreading. In such cases, the number of logical CPUs could be a multiple of the number of cores. If so, ensure you calculate the values above using the number of CPU cores and not the number of logical CPUs. If hyperthreading or dynamic threading is enabled and CPU utilization is low, you may need to change the hyperthreading configuration. See the Dynamic / Hyperthreading section below. You should only explore increasing the number of AD Parallel workers if the level of contention on all long running AD Parallel jobs is low. Similarly, you should only increase parallel_max_servers if the level of contention on all long running parallel query/DML jobs is low. Typically, this means that you should resolve SQL Tuning (poor execution plans) and contention issues before increasing the AD Parallel workers or parallel_max_servers.
Job_queue_processes
Similarly, we normally recommend job_queue_processes = number of CPU cores. If you have a large number of CPU cores (or dynamic/hyperthreading) you may want to reduce the value of job_queue_processes. However, this is only used on script adobjcmp.sql, which utilizes the job queue (DB Scheduler) to compile objects (e.g. UTL_RECOMP.RECOMP_PARALLEL).
SGA, Shared Pool, PGA etc.
If possible, you should maximize SGA and PGA sizing. You should check the feedback from AWR and the SQLT / Display Cursor (which, if statistics_level is set to ALL, should show the temporary space used by sorts, hash joins, etc). Note that the AWR advisory may underestimate the SGA required for AD Parallel jobs. For the Release 12 Upgrade, we have noticed an improvement in performance (due to reduction in physical I/O such as db file sequential read) when SGA size is increased, even though the SGA Target Advisory on AWR indicates that the
30
SGA is adequately sized. Customers have used values that are 2x the recommended value, and still seen significant benefit. We are not sure why the SGA Target Advisory does not recommend a larger value, but we suspect that it is due to the nature of the Release 12 Upgrade process (almost all rows on each table are repeatedly read; however, each AD Parallel or Parallel Execution Thread only accesses a small percentage of rows). Note that a very large SGA/buffer cache can cause excessive paging or swapping. It can also cause excessive latch waits as blocks in the buffer share the same latches. Some starting rules of thumb are: log buffer = 30 to 100 Mb shared pool = 1 to 4 GB pga target = 3 to 20 GB SGA / buffer cache = multi GB. Be generous, but avoid causing excessive paging.
db_file_multiblock_read_count
If specified, remove db_file_multiblock_read_count. This is the recommended value for normal running of Oracle E-Business Suite.
Dynamic Sampling
A number of key high volume tables (particularly in the Sub Ledger Accounting (XLA) schema) are created and populated during the Release 12 upgrade. These often have no CBO statistics gathered (prior to adsstats.sql running), and dynamic sampling could be of benefit. The Release 12 upgrade is a batch process, and most of the SQL is executed less frequently and for a larger number of rows. So increasing the dynamic sampling level should not be much of an overhead, provided you do not increase the sample size (blocks) too much. For Oracle E-Business Suite, it is normally recommended not to set the database parameter optimizer_dynamic_sampling, which means that the default value (2) will be used. However, we recommend setting a value of 4 for the upgrade, provided you revert afterwards. Note that, at level 4, the sample size is 64 blocks. Level 5, 6, 7, 8, 9 and 10 increase the sample size to 128, 256, 512, 1024, 4086 and All blocks.
31
AD Parallel Batch Size
10K is suitable for most installs: you can test other values if time allows. We do not normally recommend changing the batch size. As this is limited by the extent size on the driving table, your options to change the actual batch size used will be limited in any case. Only change the batch size if you spot serious performance issues that can be solved by increasing or decreasing the batch size (see below). And only once the SQL concerned has been confirmed as having optimal execution plans. We suggest having larger extent sizes for the larger transaction tables populated or modified by the Release 12 upgrade. It is in any case best practice to have larger extent sizes for tables with high growth. Note there are some limitations on batch size and how it is calculated. These are described below. Small extents (e.g. 128k = 16 blocks) will significantly limit the batch size. You are likely to end up with batches of much less than 1000, regardless of what batch size you select. Even if you have large extents, the actual batch sizes may vary considerably from the ones you chose because of storage overheads (percent free, block/row headers) and compression. There are advantages and disadvantages with large and small batch sizes. Larger batch sizes have less overhead, and can reduce the overhead of sub-optimal execution plans. However, there is more risk of the workers trying to perform the same operations at the same time, and thus causing contention (especially if there are multiple pieces of SQL in the job). There is also more risk of remainder batches delaying the end time of a job. Note that rows_processed column on AD_PARALLEL_UPDATE_UNITS can often contain values that do not correspond to the actual batch size used (or entered). The value of rows processed is usually SQL%ROWCOUNT following a DML statement, so it is typically the number of rows inserted or updated as a result of the SQL, not the number of rows used to drive the SQL. Sometimes there is more than one SQL/DML in the job and it is the count from the latest SQL.
Changing the AD Parallel Batch Size
Only change the batch size if you experience serious performance issues that can be resolved by increasing or decreasing the batch size. And before doing so, be sure to check that you have optimal execution plans.
32
High overhead of repeated SQL repeatedly performing the same operation => Increase batch size. High overhead of repeated full scans/range scans that cannot be solved by tuning => Increase batch size High contention between workers, occurring in highly defined peaks, at start of job (workers are doing same activity at same time/synchronized) => Decrease batch size so that workers become unsynchronized sooner. Low worker utilization at end of a job due to large remainder batch => Decrease batch size.
Be aware that changing the overall batch size is a very imprecise action. Although you may improve the performance of one job, the performance of another job may regress. You can change the batch size for individual AD Parallel jobs by hard coding the batch size in the SQL script. However, (because of the possibility of extreme values being chosen or editing errors) this is not strictly supported.
e.g.
--l_batch_size l_batch_size VARCHAR2(30) := &&2; VARCHAR2(30) := 100000;
Or in the unified driver itself. e.g.

#sql bom patch/115/sql cstmtaupg.sql none none none sqlplus_repeat &phase=upg+72 checkfile:bom:patch/115/sql:cstmtaupg.sql &un_inv &batchsize sql bom patch/115/sql cstmtaupg.sql none none none sqlplus_repeat &phase=upg+72 checkfile:bom:patch/115/sql:cstmtaupg.sql &un_inv 100000
Batch Size Calculation and its limitations
Note that (with the current version of AD_PARALLEL_UPDATES_PKG - adprupdb.pls 120.1 2006/05/08) there are some limitations on batch size and how it is calculated. This means that the actual batch sizes can often be quite different from those entered. The extent size can also cause some limitations with batch size. The code in AD_PARALLEL_UPDATE_PKG calculates the number of blocks to be used for each batch as: 1. B = (batch_size * Average_row_length)/8192
33
2. 3.
It rounds this to the nearest 10. If the average row length is 0 or the result is 0 then a default of 200 blocks is used.
Each extent on the driving table is then split into Update Units containing B blocks each, with the remainder being placed into its own update unit. Each update unit is then used as a batch. There are several limitations: If B is greater than the extent size, the whole extent is used as an update unit. Consequently, the actual batch size cannot be greater than the rows that are stored in an extent. When calculating the number of blocks (B) required for a batch, AD_PARALLEL_UPDATES_PKG does not take into account the storage overhead for the block header, percent free or row header, and so can underestimate the number of blocks required. This effect is magnified if the average row length is small. If advanced compression or compression is used, this is not reflected in the average row length. The average row length will consequently be overstated, and a larger number of blocks used for each batch. Rounding the number of blocks required to the nearest 10 adds an additional inaccuracy. Splitting each extent into update units of B blocks also adds inaccuracy. If the number of remainder blocks is much less than B, this remainder batch will be much smaller. If there is a low number of update units per extent, this will have more influence on the batch size. If the CBO Statistics are not populated, the smaller of either 200 blocks or the extent size will be used as a batch size.
This all means that the actual batch is usually smaller than that entered, and can vary in size significantly. It can be therefore be difficult to manage. Looking at the following three examples will help explain all this.
Example 1 Average Row Length 200, batch size 10000, extent size 256 blocks, percent free and storage overheads result in only 70% of block being occupied by row data.
B = 244.1 blocks = 240 blocks (rounded to the nearest 10). So that will result in two update units for each extent: one of 240 blocks, and one of 16 blocks.
34
This will in turn result in two batch sizes: 1 x 240 blocks * (8192 bytes / average row length 200) * 70% = Average of 6881 rows 1 x 16 blocks * (8192 bytes / average row length 200) * 70% = Average of 459 rows.
Example 2 Average Row Length 200, batch size 10000, extent size 16 blocks, percent free and storage overheads result in only 70% of block being occupied by row data.
B = 244.1 blocks = 240 blocks (rounded to the nearest 10). However, the extent size is 16. So the batch is limited to 16 blocks. This will result in a batch size of 16 blocks * (8192 / actual average row length 200) * 70% = Average of 459
Example 3 Average Row Length 200, batch size 10000, extent size 256 blocks, percent free and storage overheads result in only 70% of block being occupied by row data, compression ratio is 2.
B = 244.1 blocks = 240 blocks (rounded to the nearest 10). So that will result in two update units for each extent. One of 240 blocks and one of 16 blocks. This will in turn result in two batches: 1 x 240 blocks * (8192 bytes / average row length 200) * 70% * compression ratio 2 = Average of 13,762 rows 1 x 16 blocks * (8192 bytes / average row length 200) * 70% * compression ratio 2 = Average of 918 rows.
Remainder Batch Issue if Batch Size is too large
There is a risk that a very large (actual) batch size could result in a job being unduly delayed by a small number of batches that process after all other batches. Lets refer to these as remainder batches. If T = number of threads (AD workers), R = total rows to be processed, B = Batch size
35
And B is slightly less than R/T (e.g. R/(T+1)) Then this will result in T+1 batches assigned across T threads. So one thread will have twice as much work as the others. If you were processing 192,000 rows (of the driving table) on 32 workers then R/T would be around 6,000. So (assuming a large extent size) choosing a batch size of 10,000 might result in remainder batches. Therefore, in this case a much smaller batch size (such as 1,000) might be preferable. However, do not reduce the batch size excessively, as small batch sizes can result in larger overheads on each batch.
Dynamic / Hyper Threading

Many hardware vendors are now offering a version of Dynamic Threading or Hyper Threading. In these cases the hardware can be configured to present virtual CPUs (also known as Logical CPUs) to the operating system (and ultimately database). This number of virtual CPUs will be much larger than the number of cores. For example, the Oracle SPARC T4-4 has 4 sockets (physical CPUS) with 8 cores each: 32 cores in total. However, with Dynamic Threading it can run a maximum of 256 virtual CPUs (or threads). The T5-8 can have 8 T5 processors, each with 16 cores and 8 threads per core, making a total of 1024 threads. This is ideal for applications with a large number of online users, where the number of active sessions is many times the number of CPUs. Traditionally these active sessions would have shared the CPUs and there would have been a large amount of context switching. Dynamic threading can reduce context switching and thereby improve performance. However, this is not needed for cases such as the Release 12 upgrade, where there are a smaller number of active sessions, each with a high workload. In this case, the CPU capability available to each AD Parallel (or Parallel Execution) worker/thread, could be limited. Clearly a large number of workers/threads (e.g. 256) all performing the same operation will result in massive contention that will far outweigh the benefits of greater CPU utilization. So the recommendation is to configure the dynamic threading/hyper threading for maximum throughput per CPU cycle i.e. 1 thread for each core. On Oracle SPARC T4-4 you can do this by setting dynamic threading to max-ipc.
36
In addition, you will need to disable any power management as this will either prevent configuration of threading for maximum throughput, or throttle the CPU capacity.
Gathering CBO Statistics

See My Oracle Support Knowledge Document "EBPERF FAQ - Collecting Statistics with Oracle EBS 11i and R12 (Doc ID 368252.1)".
Pre-install Patch 16410424.R12.FND.B before R12.1.1
Patch 16410424.R12.FND.B should be pre-installed before R12.1.1 This does not have any particular advantage for the upgrade itself. But it could help post upgrade. It properly enables auto sample size (dbms_stats.auto_sample_size) in fnd_stats for 11G. Although you can enable this on previous versions of fnd_stats simply by setting estimate_percent to dbms_stats.auto_sample_size (=0) when you call fnd_stats.gather_schema_statistics or fnd_stats.gather_table_stats. It enables extended column statistics in fnd_stats.
Gathering Oracle E-Business Suite Schema Statistics
In the pre-upgrade environment, you should have gathered schema statistics for all Oracle E-Business Suite schemas using FND_STATS (or the Gather Statistics concurrent program) with the GATHER_AUTO option. The Release 12 upgrade will gather all Oracle E-Business Suite schema statistics again (using adsstats.sql) towards the end of the R12.1.1 upgrade (last+63). Note that adsstats.sql is not run in the R12.1.3 RUP. It is important to be aware of the difference between adsstats.sql and adstats.sql. They are two distinct scripts: adsstats.sql gathers Oracle E-Business Suite schema statistics. adstats.sql disables the 10g/11g automatic statistics gather jobs, and gathers the Dictionary and Fixed Object Statistics.
After the Release 12 upgrade, but prior to the system being available to users, you should gather CBO statistics again for all Oracle E-Business Suite schemas using FND_STATS (or the Gather Statistics concurrent program) with the GATHER_AUTO option. Do not use DBMS_STATS or Analyze.
37
In principle, you should not need to gather CBO statistics for Oracle E-Business Suite schemas at any other time during the Release 12 upgrade. However, in practice you may need to gather or delete statistics for specific objects as a result of performance issues (see Resolving Performance Issues section), especially on tables created by the Release 12 upgrade or on temporary/transient tables.
Importing Oracle E-Business Suite Schema Statistics
If the adsstats.sql script is taking a significant amount of time to run, you could reduce the Release 12 upgrade time by: Exporting statistics gathered during test runs (by adsstats.sql - at same point: phase: last+63). Importing these statistics instead of running adsstats.sql.
However, this does complicate the upgrade. So only do this if you estimate that the time saved will be significant, and you are confident in adopting this approach. Note that if parallel_max_servers is set to a value much less than 2 x number of cores, you should consider increasing parallel_max_servers before resorting to export/import. If importing statistics is also long running, it may have a performance issue that could be resolved by gathering fixed object and dictionary statistics prior to importing. You would normally only gather fixed object and dictionary statistics after the Release 12 upgrade is complete (i.e. after R12.1.3 RUP). So these are additional activities and their elapsed time should be deducted from any gains made by importing statistics. You can export the statistics using: FND_STATS.backup_schema_stats(schema_name => ALL, statid => '<your statid>')
And import them using:
FND_STATS.restore_schema_stats(schema_name => ALL, statid => '<your statid>') There are two reasons you should use FND_STATS.backup_schema_stats and restore_schema_stats: Only FND_STATS is supported with Oracle E-Business Suite. FND_STATS.backup_schema_stats and restore_schema_stats will only backup and restore Oracle E-Business Suite schemas (around 85% of objects on an Release 12 database). In contrast, DBMS_STATS.export_database_stats and import_database_stats will export/import
38
all schema statistics and also gather fixed object, dictionary and system statistics. So the workload will be considerably less if you use FND_STATS. We suggest using the following SQL scripts for exporting and importing statistics.
Exporting Statistics set verify off whenever sqlerror exit failure rollback; whenever oserror exit failure rollback; declare begin FND_STATS.BACKUP_SCHEMA_STATS(schemaname => 'ALL', statid => '<your identifier>'); exception when others then raise_application_error(-20000, sqlerrm ||' Error while executing FND_STATS.BACKUP_SCHEMA_STATS package.'); end; / exit;
Importing Statistics set verify off whenever sqlerror exit failure rollback; whenever oserror exit failure rollback; declare begin FND_STATS.RESTORE_SCHEMA_STATS(schemaname => 'ALL', statid => '<your identifier>'); exception when others then raise_application_error(-20000, sqlerrm ||' Error while executing FND_STATS.RESTORE_SCHEMA_STATS package.'); end; / exit;
However, at the time of writing there was an issue with DBMS_STATS.IMPORT_SCHEMA_STATS: if a schema has no tables, an exception will be raised that causes FND_STATS.RESTORE_SCHEMA_STATS to fail before completion). There are several Oracle EBusiness Suite schemas with no tables. A workaround is to use the following script:
set verify off whenever sqlerror exit failure rollback; whenever oserror exit failure rollback; DECLARE
39
CURSOR schema_cur IS WITH schema_tabcount AS (SELECT c5, COUNT(*) num_obj FROM fnd_stattab WHERE statid = <your identifier>' GROUP BY c5 ) SELECT upper(oracle_username) sname FROM fnd_oracle_userid WHERE oracle_id BETWEEN 900 AND 999 AND read_only_flag = 'U' AND EXISTS (SELECT 'exists' from schema_tabcount WHERE c5 = upper(oracle_username) AND num_obj >0) UNION ALL SELECT DISTINCT upper(oracle_username) sname FROM fnd_oracle_userid a, fnd_product_installations b WHERE a.oracle_id = b.oracle_id AND EXISTS (SELECT 'exists' from schema_tabcount WHERE c5 = upper(oracle_username) AND num_obj >0) ORDER BY sname; BEGIN FOR c_schema IN schema_cur LOOP BEGIN DBMS_STATS.IMPORT_SCHEMA_STATS(c_schema.sname, 'FND_STATTAB', <your identifier>', 'APPLSYS'); EXCEPTION WHEN OTHERS THEN dbms_output.put_line('Error on '||c_schema.sname); dbms_output.put_line(SUBSTR(SQLERRM,1,80)); END; END LOOP; EXCEPTION WHEN OTHERS THEN raise_application_error(-20000, sqlerrm ||' Error while executing adsstats_restore'); end; / exit;
Further strategies for reducing adsstats.sql elapsed time
If the adsstats.sql job is taking a long time during the R12.1.1 upgrade there are two further approaches that may help for large environments, where there are tables with more than 100 million rows.
Use dbms_stats.auto_sample_size in adsstats.sql
If on 11g then amend the adsstats.sql job so that it calls FND_STATS.GATHER_SCHEMA_STATISTICS with an estimate_percent of
40
dbms_stats.auto_sample_size (=0) rather than 10%. This now uses hash-based sampling that has accuracy close to 100% but executes much more quickly. For larger tables (e.g. 10-100 million rows) it will normally be a little quicker than an estimate percent of 10%. For higher volumes it could be much quicker. However, for smaller tables it could be slower. So this will only be effective on large environments, with 100 millions of rows on the new or updated tables for R12 (e.g. XLA schema, ZX schema, AP_DBI_LOG, AP_LIABILITY_BALANCE, AP_INVOICE_DISTRIBUTIONS_ALL). For smaller environments it could actually increase the time taken.
Sample specific long running tables at a lower percentage
You can check the table FND_STATS_HIST to see which tables are taking the longest. Use the following SQL to show the latest FND_STATS runs (request_id) with the number of tables analyzed. You can then identify the <request_id> for the adsstats.sql run in the R12 upgrade.
SELECT request_id, count(*) tables_analyzed, TO_CHAR(MAX(last_gather_end_time),'DD-MON-YYYY HH24:MI:SS') max_end_time, MAX(last_gather_end_time) max_end_time_int FROM fnd_stats_hist WHERE object_type = 'CASCADE' AND last_gather_end_time IS NOT NULL GROUP BY request_id ORDER BY MAX(last_gather_end_time) DESC
Once you have identified the request_id, the following SQL will give you the tables where gathering stats took the longest along with information on the number of rows, blocks and sample size.
SELECT fsh.schema_name, fsh.object_name, fsh.object_type, ROUND((fsh.last_gather_end_time - fsh.last_gather_start_time)*24*3600,0) time_secs, to_char(fsh.last_gather_start_time,'DD-MON-YYYY HH24:MI:SS') start_time, to_char(fsh.last_gather_end_time,'DD-MON-YYYY HH24:MI:SS') end_time, dt.num_rows, dt.blocks, dt.avg_row_len, ROUND(DECODE(NVL(num_rows,0),0,NULL,(sample_size*100)/num_rows),0) est_pct, sample_size FROM fnd_stats_hist fsh ,dba_tables dt WHERE dt.owner = fsh.schema_name AND dt.table_name = fsh.object_name AND fsh.request_id = <request_id> AND fsh.object_type = 'CASCADE'
41
AND fsh.last_gather_end_time IS NOT NULL ORDER BY (fsh.last_gather_end_time - fsh.last_gather_start_time)*24*3600 desc
The longest running ones are candidates for running at a lower estimate percentage (sample size) on the next run. By sampling at a lower percentage you will be reducing the accuracy of the CBO statistics, so take care to monitor the performance of post R12.1.1 activities (e.g. R12.1.3 driver and post upgrade patches/jobs). You can then gather these tables at a lower percentage directly before running adsstats.sql. Or you can edit adsstats.sql directly to gather these tables at a lower percentage prior to gathering schema stats (FND_STATS.GATHER_SCHEMA_STATISTICS). The command is:
fnd_stats.gather_table_stats('<owner>', <table_name>', <estimate_percent>, <parallel_degree>);
Where <parallel-degree> is the setting of parallel_max_servers (e.g. v_parallel in adsstats.sql itself). It is advised that you use: 3 percent for tables between 100 million and 1 billion rows 1 percent for tables with more than 1 billion rows.
Incremental Statistics for Partitioned Tables
Incremental statistics gathering is a new Oracle 11gR2 DBMS_STATS feature and is fully supported by FND_STATS. Oracle derives global statistics by scanning only new or modified partitions rather than the whole table, which is highly beneficial with high volume Oracle E-Business Suite tables. This will have no impact for partitioned tables during the upgrade itself, but it will speed up statistics gathering post go live. It should be set for the following new partitioned tables: XLA_AE_HEADERS, XLA_AE_LINES, XLA_DISTRIBUTION_LINKS, XLA_EVENTS, XLA_TRANSACTION_ENTITIES, XLA_TRIAL_BALANCES, AP_DBI_LOG, AP_LIABILITY_BALANCE. You run the following command to invoke it: dbms_stats.set_table_prefs(<schema_name>, '<table_name>',
Locked Statistics
'INCREMENTAL', 'TRUE');
42
You may have had reason to lock the statistics on some tables. For example, the table could be a temporary or transient table (or Global Temporary table), where you have gathered a representative set of statistics that you wish always to be used and never overwritten. In such cases, be aware that the adsstats.sql script will unlock the statistics and re-gather. And it is likely that these tables will be empty (or not have representative data volumes) during the Release 12 upgrade. It would be useful to obtain a list of these prior to the upgrade, so that they can be locked and excluded. For Oracle E-Business Suite you should lock statistics using the DBMS_STATS.LOCK_TABLE_STATS procedure, and then exclude them from subsequent FND_STATS gathering jobs by using FND_STATS.LOAD_XCLUD_TAB procedure. Once the statistics are locked using the DBMS_STATS.LOCK_TABLE_STATS procedure, FND_STATS will skip the tables on which statistics are locked in the subsequent statistics gathering jobs. Gather Schema / Table Statistics Concurrent Program will display a message in the request log file saying that statistics are locked on the table. FND_STATS.LOAD_XCLUD_TAB will tell FND_STATS to skip the table from the Gather Schema Statistics Concurrent Program. It will seed an entry in the FND_EXCLUDE_TABLE_STATS table.
These should have been gathered on the pre-upgrade environment. Oracle Database 10g and above use the CBO for most of the dictionary SQL, so you must gather statistics on the dictionary objects as well as the fixed objects. A fixed object (X$ tables) resides in memory only, and typically records the information about the instance or memory structures. The v$ dynamic performance views are defined on top of X$ tables e.g. V$SQL and V$SQL_PLAN. Data dictionary tables (e.g. SYS.USER$, SYS.TS$, SYS.SEG$, SYS.OBJ$, SYS.TAB$, SYS.FILE) are stored on data files like normal application tables. If you see internal SQL (on V$ views or on SYS/SYSTEM objects) appearing high in AWR and TKPROF reports, it is likely that you need to gather dictionary and fixed object statistics. Note that the FND_STATS API does not gather statistics for dictionary or fixed objects. You need to use the DBMS_STATS APIs.
43
Fixed Object Statistics
Typically, you will need to gather fixed object statistics when there have been significant changes to fixed objects. For example, after: A major database upgrade. Platform upgrades, especially Exadata. Application upgrades, or addition of a new application module. Changes to the SGA/PGA configuration of the database. Significant changes in the workload or number of sessions.
For optimal performance, statistics on fixed objects should be gathered when there is representative activity (typical load) on the system. The command to run is: execute DBMS_STATS.GATHER_FIXED_OBJECTS_STATS; For the Release 12 upgrade, the fixed object statistics should be gathered:
After any associated platform or database upgrade that is part of the overall Oracle E-Business Suite upgrade. After any SGA/PGA parameters have changed. After the Release 12 upgrade, when there is representative activity on the system.
Dictionary Statistics
Normally, the Dictionary Statistics need to be analyzed after a sufficient number of DDL operations have occurred in the database (for example, if there is a significant change in the database when a lot of new objects are created or rebuilt). Use the following command (in Oracle Database 10g and above) to gather statistics for all system schemas, including SYS and SYSTEM. execute DBMS_STATS.GATHER_DICTIONARY_STATS( estimate_percent => DBMS_STATS.AUTO_SAMPLE_SIZE, options => 'GATHER AUTO');
44
For Oracle Database 9i, the command is: execute DBMS_STATS.GATHER_SCHEMA_STATS(''SYS'', method_opt=>''for all columns size 1'', degree=>30, estimate_percent=>100, cascade=>true); For the Release 12 upgrade, Dictionary Statistics should be gathered: After any associated platform or DB upgrade that is part of the overall Oracle E-Business Suite upgrade. After the Release 12 upgrade. After move to OATM.
adstats.sql
This script: Disables the Oracle 10g/11g Automatic Statistics gather jobs. Gathers Dictionary and Fixed Object Statistics.
However, it can only be run on Database 10g or above, and with the database in restricted mode. Do not use this script to gather fixed object or dictionary statistics as part of the Oracle E-Business Suite Release 12 upgrade itself (although it may be run as part of an associated database upgrade). Use the APIs directly, as instructed above. It is recommended that this script is run as part of the upgrades of Oracle E-Business Suite instances to Oracle Database 10g and 11g. Details are included in the upgrade instructions. Notes: The adstats.sql script is not run automatically as part of the Release 12 upgrade, and will not be in the R12.1.1, R12.1.3 unified drivers. For optimal performance, statistics on fixed objects should be gathered when there is a typical load in the system. This script is designed to be run only when the database is in restricted mode. The script does not gather system statistics. As mentioned earlier, adstats.sql is sometimes confused with adsstats.sql, which gathers Oracle E-Business Suite schema statistics.
45
Resolving Performance Issues

This section includes guidance on: Diagnostics to gather during the upgrade, particularly diagnostics to gather when you encounter long running jobs and SQL. Common (generic) issues and their solutions. Specific issues will be covered in bugs on My Oracle Support.
Diagnostics to be gathered during the upgrade

For most upgrades, we recommend that Automatic Workload Repository is enabled, with a snapshot of 30 minutes. For shorter upgrades, a smaller snapshot may be more suitable, allowing analysis of shorter time periods. The AWR retention period should be long enough to cover the duration of the upgrade run plus a significant period afterwards (to gather diagnostics and analyze). The suggestion is N+7 days, where N is the estimated upgrade time, but a longer period will provide more time to gather subsequent diagnostics and statistics. We also strongly recommend that statistics_level is set to ALL for the duration of the Release 12 upgrade dry runs.
Express Diagnosis of Oracle E-Business Suite Release 12 Upgrade Performance Issues
When analyzing Release 12 upgrade performance issues, your goal is to: Prevent wasted test iterations. You should aim to provide solutions that solve the issue first time. Maximize the number of performance issues resolved.
Upgrade jobs cannot be tested in isolation. They can only be tested on the next iteration. If a fix does not work, it is a potential wasted test iteration. To do this you need: Actual statistics: So that you can see exactly which execution plan steps are inefficient, rather than those that you think might be inefficient. You can also estimate the likely impact of a performance fix. There will then be a higher probability of providing a fix that can solve the
46
performance issue first time. Also, you will be able to identify marginal fixes (i.e. fixes that reduce elapsed times by 10-50%, for example by having more specific index access). These fixes often reduce contention between workers. Diagnostics that are quick and easy to run, so that you can quickly and easily obtain diagnostics on jobs and SQL. Diagnostics that have very little impact on the performance of the Release 12 upgrade. If they can safely be run during the upgrade; you obtain the results sooner and may be able to resolve the issue more quickly. Before the Release 12 upgrade, set the statistics level to ALL Identify top SQL in AWR (see useful scripts) Obtain Display Cursor Reports (ALL +ALLSTATS) for long-running SQL Obtain SQL Monitor Reports for SQL that uses Parallel Query or DML Identify when a piece of SQL ran (see useful scripts) Match long-running SQL with a job (see useful scripts) Report on CBO Statistics for all Oracle E-Business Suite tables (see useful scripts) Obtain AD Job Timing Reports Identify long-running upgrade Jobs (see useful scripts) Obtain the file versions for long running jobs Obtain AWR reports Run the afxplain.sql script to obtain detailed CBO statistics, database parameters, CBO parameters, and metadata: do this for individual pieces of SQL.
The key preparatory step is:
During the upgrade:
After the Release 12 upgrade:
Other Diagnostics
For more detailed analysis you will require the following:
47
AD utility and worker logs. SQLT With XTRACT for long-running SQL AWR Export AD Parallel tables export
In addition, Oracle Support and Development may require:
See the Diagnostics During Upgrade section in Resolving Performance Issues below for more details.
Statistics_Level = ALL
This can be set using the command: SQL>alter session set statistics_level='ALL' This is the simplest way to see actual execution statistics (including elapsed time, physical reads, buffer gets etc) for each execution plan line (on SQLT and Display Cursor report). The alternative of SQL Trace and TKPROF requires editing standard code. Using this strategy will typically speed up the resolution of issues significantly. and may also allow you to identify the correct solution first time. Many technical architects and DBAs at customers (or implementing partners) can be resistant to setting statistics_level = ALL, believing that this can slow down performance significantly. Two points are relevant here: Although setting statistics_level = ALL will have some performance impact, it is likely to be small and not significant. The Release 12 upgrade is made up of batch processes, and so the statistics workload is a much lower proportion of the total. Even if the performance impact is significant, the goal is to reduce the elapsed times for the latter dry runs and go live (when it will be feasible to revert statistics_level to its previous value). So suffering an increase in elapsed time during an early stage of testing is not an issue.
So there may be a small impact on elapsed time and the work you have to do initially, but it will help you subsequently reduce the elapsed time and amount of re-work you have to do.
48
Note that setting statistics_level to ALL while AWR is enabled could significantly increase the number of rows inserted to the WRH$_LATCH_CHILDREN table in AWR. So you should monitor the SYSAUX tablespace to ensure that it does not run out of space.
SQL Trace SQL Script Specific
If you are unable to obtain the diagnostics that would enable you to resolve a performance issue from the display cursor, SQL monitor and AWR, then you should obtain a 10046 SQL Trace (with waits / level 8) for the script/job on the next test run. You can enable SQL Trace at system level. However, this will produce a large number of trace files, and it may be difficult to locate the correct traces. (Tip: on UNIX, you can use grep to search files for key SQL strings or script/module names). It is advisable to enable SQL Trace for specific scripts by temporarily editing the SQL script as follows: Add the following before the main body of the script. alter session set tracefile_identifier=<identifier string>; alter session set events '10046 trace name context forever, level 8; alter session set statistics_level=ALL ; alter session set max_dump_file_size=UNLIMITED; alter session set events '10046 trace name context off';
Add the following after the main body of the script.
Another advantage of enabling the trace at script level is that you can give each script different trace identifiers, so finding the traces becomes easy.
SQL Specific
Once you have used AWR or TKPROF to identify the long-running SQL (in the long-running jobs), you should provide the output below: Note that for these to be useful, statistics_level should be set to ALL. You need the actual statistics to be able to identify the expensive execution plan steps or execution plans that are moderately sub-optimal. (The execution plan itself may look OK without this information).
49
Display Cursor Report
This displays the actual execution plan of any cursor loaded in the cursor cache. At the basic level it shows the actual execution plan. However, the format ALL also includes extra information such as pruning, parallel execution, predicate, projection, alias and remote SQL information. The +ALLSTATS option will include actual statistics for each execution plan step. ALLSTATS includes IOSTATS and MEMSTATS. For example: Elapsed time Physical reads Buffer gets Memory used (in PGA) for memory intensive operations (such as hash-joins, sorts, bitmap operators etc).
However, this is only provided if statistics_level=ALL Most of the information is also included in SQLT with XTRACT method. However, display_cursor provides a simpler view of the information. It can also be run during the upgrade, while the long running SQL is in progress, without much of an overhead. Note that you should run the display cursor report as soon as possible. If it is delayed, the cursor may have been flushed from memory or invalidated. In such a case, no data will be available. The report can be produced by running the following SQL script:
SET pages 0 SET lines 300 SET LONG 10000 SET LONGCHUNKSIZE 10000 SPOOL<report_name>.txt SELECT * FROM TABLE(dbms_xplan.display_cursor('<sql_id>', NULL, 'ALL +ALLSTATS')); SPOOL OFF;
If the SQL is no longer in memory, but is in AWR, you could use the Display AWR report (DBMS_XPLAN.DISPLAY_AWR) instead. However, this does not report on actuals: it does not have a +ALLSTATS option, and there are no actual statistics for execution plan steps stored in AWR. Note that the display cursor and AWR reports only show the sql_text (first 1000 characters) and not the full_text.
50
So you may need to run the following SQL script to obtain the full SQL text (if you do not have it already).
SET pages 0 SET lines 300 SET LONG 10000 SET LONGCHUNKSIZE 10000 SPOOL<report_name>.txt SELECT sql_id, sql_text, sql_fulltext FROM v$SQL WHERE sql_id = '<sql_id>'; SPOOL OFF; SQL Monitor Report
The main advantage of this is that it gives a good view of how parallel SQL/DML performs across stages of the plan and parallel slaves. However, it is only available in Oracle Database 11g. It can be run during the upgrade, while the long running SQL is in progress, without much of an overhead. You can produce it by running the following SQL script:
set trimspool on set trim on set pages 0 set long 10000000 set longchunksize 10000000 set linesize 200 set termout off spool sql_monitor_for_<sql_id>.htm variable my_rept CLOB; BEGIN :my_rept := dbms_sqltune.report_sql_monitor(sql_id => '<sql_id>', report_level => 'ALL', type => 'HTML'); END; / print :my_rept spool off; set termout on
Again, statistics_level must be set to ALL.

afxplain.sql
This script is found in $FND_TOP/sql, and Provides detailed CBO statistics, database parameters, CBO parameters, and metadata for SQL. It can be run during the upgrade, as it does not have much overhead.
51
Create a file containing the SQL you want the diagnostics for. In this example it is called query.sql. Run: afxplain.sql query.sql Y N The Y N parameters respectively specify that you want a CBO Trace (10053), but you do not want to export statistics. The output will be in file query.sql.out The trace for the 10053 event will be in the user dump destination.
Note that the output will not contain the following:
Actual counts of rows on each table. Actual execution plan, but you should have this from the display_cursor (or display_awr) report above.
SQLT Output using the XTRACT method
This uses a significant amount of system resources, so should not be run during the upgrade. Instead, you can use Display Cursor Report and afxplain.sql to obtain much of the information that is included in the SQLT output. See SQLT (SQLTXPLAIN) - Tool that helps to diagnose a SQL statement performing poorly or one that produces wrong results (Doc ID 215187.1) The SQLT should be provided on the same environment where the performance issue was observed, and should be run as soon after the relevant program/process as possible. Be aware of any actions that may alter the data that SQLT is capturing (that is, actions that take place after the observed performance issue, but before SQLT is run). For example, statistics being gathered, removed, locked, imported, or data being deleted from temporary or interface tables. To run SQL with XTRACT you will need the sql_id or hash_value. In Oracle Database 11g, the TKPROF will give the sql id and hash value. For Oracle 11g. the raw trace gives the sql_id (sqlid=). In 10g and before, it gives the hash value (hv=). Other methods of obtaining the sql_id include using AWR, Oracle Enterprise Manager, and directly querying V$SQL.
52
The SQLT provides execution plans, CBO statistics, database parameters, CBO parameters, performance statistics, and metadata (schema object definitions and so on). It is particularly useful if you do not have access to the instance. Even if you do have access, it gives you all the information in one place. You need XTRACT because it gives the actual execution plans. If statistics_level = ALL then you get the actual I/O, buffer gets and elapsed time. It is also Oracle RAC-aware. Note that SQLT runs AWR and ASH reports. Some dictionary objects (particularly WRH$_LATCH_CHILDREN) will have grown significantly during the upgrade. So you should gather fixed object and dictionary statistics before running SQLT, especially if you have set statistics level to ALL (and you normally do not).SQLT can take quite a while to run. To reduce the workload, it is recommended that you run the following (from SQL*Plus) before running sqltxtract.sql: To disable Test Case Builder TCB and/or SQL Tuning Advisor EXEC sqltxplain.sqlt$a.set_param('test_case_builder', 'N'); EXEC sqltxplain.sqlt$a.set_param('sta_time_limit_secs', '30'); To disable the automatic export of a test case repository EXEC sqltxplain.sqlt$a.set_param('export_repository', 'N'); If SQLT still takes a long time, and the schema objects behind your SQL contain a large number of
sub-partitions, you can reduce the granularity of the data collected:
EXEC sqltxplain.sqlt$a.set_param('c_gran_segm', 'PARTITION'); EXEC sqltxplain.sqlt$a.set_param('c_gran_cols', 'PARTITION'); EXEC sqltxplain.sqlt$a.set_param('c_gran_hgrm', 'PARTITION'); Note that these commands can all be run as APPS. They do not need to be run as user SQLTXPLAIN These values are stored in a table called SQLTXPLAIN.SQLI$_PARAMETER. Once they are set, they do not need to be re-set for each execution of SQLT. You can check the current values of these parameters by querying the table SQLTXPLAIN.SQLI$_PARAMETER. You can disable the counting of tables, and reduce the time further, by running the following. However, you will lose information on the actual number of rows in each table. EXEC sqltxplain.sqlt$a.set_param('count_star_threshold', '0');
53
All of this assumes that you are using a SQLT version greater than 1.4.4.4 (April 2, 2012)
AWR Export
The AWR can be exported to a dump file (data pump export file) using $ORACLE_HOME/rdbms/admin/awrextr.sql. The $ORACLE_HOME/rdbms/admin/awrload.sql script can be used to load the export dump into a local database, where you can then run SQL against various AWR tables to get detailed analysis of where waits occur (job, SQL, object etc), and any patterns. The key tables are DBA_HIST_ACTIVE_SESS_HISTORY, DBA_HIST_SEG_STAT_OBJ, DBA_HIST_SYSTEM_EVENT , DBA_HIST_SEG_STAT, DBA_HIST_WAITSTAT, DBA_HIST_ACTIVE_SESS_HISTORY, DBA_HIST_SQLSTAT, DBA_HIST_SQLTEXT. Alternatively, you can access the source environment and run SQL directly (i.e. without the need to run export/import). One advantage of the export /import strategy is that you can keep the diagnostics after the original environment has been refreshed. See Oracle Database Performance Tuning Guide 11g Release 2, Transporting Automatic Workload Repository Data (p5-19)
AWR Reports
You should obtain AWR reports for the period that the upgrade is running. You should also obtain AWR reports for the duration of long running jobs (i.e. for the time between the snapshots taken just before a job starts and just after it finishes). Typically, the snapshot interval should be 30 minutes (the default is 1 hour). This can be altered to suit the overall elapsed time of your upgrade/window. If you have a shorter upgrade/window, then smaller delays are significant, and you may want a shorter snapshot. For upgrades taking 8-15 hours, a snapshot of 15 minutes would be more suitable. When running AWR reports for specific jobs you want the report to include the period that the job was running, without containing too much before or afterwards. You will typically use awrrpt.sql to generate the AWR reports. Always choose HTML report type. On an Oracle RAC instance, awrrpti.sql will usually suffice, as you will be running the upgrade on one Oracle RAC node only.
54
The three main starting points to look for in AWR reports are long-running SQL (SQL Statistics /SQL Ordered By); contention/bottlenecks (Top 5 Timed Foreground Events/Foreground Wait Events); and CPU utilization. However, many other useful statistics can be obtained from AWR reports, depending on the situation. Some are covered elsewhere in this document. If you see high levels of a particular wait(s), you should first check to see if it only occurs with particular long-running SQL and jobs, before you assume it is a system configuration or resource issue. You can also run: awrsqrpt.sql (or awrsqrpi.sql) to report on the execution plans for particular long-running SQL. ashrpt.sql (or ashrpti.sql) to report on the Active Session History for particular SQL. This gives useful information on the waits and events for each row source or object.
Note that some dictionary objects (particularly WRH$_LATCH_CHILDREN) and fixed objects will have grown significantly during the upgrade. This is especially the case if you have set statistics level to ALL, or you have changed to a high retention period or a short snapshot interval. So you should gather fixed object and dictionary statistics before running AWRs,
AD Parallel tables export
The key tables to export (either using Data Pump Export or Original Export) are: AD_PARALLEL_UPDATES AD_PARALLEL_UPDATE_UNITS AD_TASK_TIMING
These can then be imported onto a local database using either Data Pump Import or Original Import, and then be analyzed. The key items you are looking for here are the actual rows processed on each batch; the number of batches; the progress over time (to spot jobs that slow down); and any significant irregularity in rows processed/time taken for each batch that might suggest locks/sleeps or data distribution issues. Also the start times and end times of jobs and workers (AD_TASK_TIMING), which can help us match long running SQL and performance issues (on AWR etc) with specific jobs. The AD Job Timing Report (below) only reports the Top 100 Time Consuming Jobs. However, for AD Parallel jobs each worker is reported as a separate job, so if you have specified multiple workers
55
(for example, 16, 24 or 32), it may only report a handful of jobs. In this case, you can query the AD_TASK_TIMING table to identify all the long running jobs. You can also query AD_TASK_TIMING to identify low worker utilization (and potential resource utilization) due to phasing waits. See the Useful Scripts part of this document for report that will show all long running jobs and phasing waits.
AD Job Timing Report
This reports: Number of successful, failed, deferred, re-started or skipped jobs. The top 100 time consuming jobs. The failed, deferred, re-started and skipped jobs. The timing of each upgrade phase, with the total number of jobs, and the number deferred, re-started and skipped.
However, it only reports the Top 100 Time Consuming Jobs, and for AD Parallel jobs it considers each worker to be a different job. This means it may only report a handful of jobs, and you will need to query the AD_TASK_TIMING table directly (see section above and Useful Scripts section). When you run AutoPatch or AD Administration, it automatically generates an AD Job Timing report (adt<session_id>.lst). You can access the contents of this report from Oracle Application Manager, or obtain reports for completed upgrade sessions from the APPL_TOP/admin/<SID>/out directory. The report is called adt<session_id>.lst. You can also run the AD Job Timing Report for AD Administration jobs from the command line. $ cd $APPL_TOP/admin/<SID>/out $ sqlplus <APPS username>/<APPS password> @$AD_TOP/admin/sql/adtimrpt.sql \ <session id> <output file> Where <session_id> is the session of the timing statistics you want to see, and <output file> is the name of the file where the statistics will be written.
56
You can also run $AD_TOP/admin/sql/adtimdet.sql in a similar way. This gives details on all jobs ordered by phase or elapsed time. This is useful if you want to find out how long any job took to run, and also where the Top 100 Time Consuming Jobs is dominated by multiple workers of a few jobs. Note that the SQL scripts may be in $AD_TOP/sql (not admin/sql).
AD Utility and Worker Logs
All AD utilities record their processing actions and any errors that they encounter in log files. Many utilities prompt you for the name of the log file, but default to <utility_name>.log. For example, for AD Administration the default log file is adadmin.log. For AutoPatch, it is adpatch.log. AD utilities that process jobs in parallel also write details to worker log files. The adwork<number>.log files (adwork001.log, adwork002.log...) reside in the $APPL_TOP/admin/<SID>/log directory, where <SID> is the value of the ORACLE_SID or TWO_TASK variable (UNIX).
Long Running SQL, Contention and Tuning
For long running jobs or SQL, it is best to start by you should first investigating if good execution plans are being used. A poor execution plan (or even just one that is moderately sub-optimal) can be the root cause of contention, especially if that contention only occurs during a particular job. Often a sub-optimal execution plan can cause contention and/or additional waits because it is accessing unnecessary database blocks. This might be because of a full table scan, or an index that is not particularly selective. This increases the probability of different AD workers accessing the same blocks at the same time. Once the execution plan no longer accesses those unnecessary blocks, the contention/waits may reduce or disappear. However, to identify sub-optimal execution plans, you need to check the actual row counts, buffer gets, and elapsed time of each execution plan step. Quite often execution plans that look good at first glance (e.g. use selective indexes) may not be close to the best. Also note that an execution plan that is quite good may still have one step where some unnecessary blocks are accessed, increasing the risk of contention as you add workers or threads. This is why we advise setting statistics_level = ALL.
File Versions
When resolving specific issues, Oracle Support and Development will ask for the versions of files used in the upgrade.
57
For your own purposes, you may want to check what is done in a particular job, or the exact metadata/object definition in a file. When doing this, be sure to identify the version of the files used in the upgrade. For example, for an R12.1.3 upgrade, the version present after all the upgrade steps including R12.1.3 and post upgrade steps may be different from the one used during the R12.1.1 upgrade. To find out the version of a file used during the R12.1.1 upgrade, check the unified driver used for that part of the upgrade: the same principle applies to the 12.1.3 upgrade. (For example: $ cat u_merged.drv|grep A5 cstpostimportaad.sql).
A Little Contention is good
Once you have removed any unnecessary contention caused by sub-optimal execution plans, a small amount of contention (e.g. 5 to 15% on particular waits) between AD Parallel or Parallel Execution sessions can be a useful indicator that you are getting the most out of the available resources. However, measuring the level of contention for some events is difficult. For typical contention waits due to sessions accessing the same block (in buffer), latch or row, you can consider the whole of the wait as contention. But for waits on resources such as disk (e.g. db file sequential read), you have a minimum average wait time even when the resource is not busy. In such a case, you should be looking at how long the wait has increased, and the length/fatness of the tail on the histogram.
Shifting Bottlenecks
Be aware of shifting bottlenecks. Sometimes you take an action that reduces a particular wait significantly, but another wait significantly increases. The obvious assumption is that the action has caused the new wait. However, this may not be the case: it may just have revealed the wait. This does not mean that you shouldnt take action to remove the new wait. It just means that reversing the original action may not be the solution. Here is an illustration of this effect.
58
Sessions arrive at first event in random pattern and they have to wait. So the first event is a bottleneck. Because the first event/bottleneck is in effect regulating the flow, there are few waits at the second event.
When the first bottleneck is removed, the sessions arrive at the second event in a random pattern and have to wait. The second event then becomes a bottleneck.
Useful Scripts
Top SQL between two snapshots in AWR
SELECT * FROM (SELECT dhs.sql_id, ROUND(SUM(dhs.elapsed_time_delta/1000000),0) elapsed_time_secs, ROUND(SUM(dhs.cpu_time_delta/1000000),0) cpu_time_secs, SUM(dhs.disk_reads_delta) disk_reads, SUM(dhs.buffer_gets_delta) buffer_gets, SUM(dhs.px_servers_execs_delta) px_server_execs, SUM(dhs.rows_processed_delta) rows_processed, SUM(dhs.executions_delta) executions, ROUND(SUM(dhs.iowait_delta/1000000),0) iowait_secs, ROUND(SUM(dhs.clwait_delta/1000000),0) clwait_secs, ROUND(SUM(dhs.ccwait_delta/1000000),0) ccwait_secs,
59
ROUND(SUM(dhs.apwait_delta/1000000),0) apwait_secs FROM dba_hist_sqlstat dhs ,v$database d WHERE dhs.dbid = d.dbid AND snap_id > <begin snap> and snap_id <= <end snap> GROUP BY dhs.sql_id ORDER BY 2 DESC) WHERE ROWNUM <= 100;
Where <begin snap> and <end snap> are the start and end snapshot IDs.
Identify when a piece of SQL ran
The following SQL will show when a particular piece of SQL ran (i.e. between which snapshots). This is useful in matching SQLs to jobs.
SELECT dhs.sql_id, dsn.snap_id, dsn.begin_interval_time, dsn.end_interval_time, ROUND(SUM(dhs.elapsed_time_delta/1000000),0) elapsed_time_secs FROM dba_hist_sqlstat dhs ,v$database d ,dba_hist_snapshot dsn WHERE dhs.dbid = d.dbid AND dsn.snap_id = dhs.snap_id AND dsn.dbid = dhs.dbid AND dsn.instance_number = dhs.instance_number AND dhs.sql_id = '<sql_id>' AND dsn.snap_id > <begin_snap> and dsn.snap_id <= <end_snap> GROUP BY dhs.sql_id, dsn.snap_id, dsn.begin_interval_time, dsn.end_interval_time ORDER BY dsn.snap_id;
Where <begin snap> and <end snap> are the start and end snapshot IDs.
Long Running Upgrade Jobs
Note that the Top 100 Time Consuming Jobs section of the standard adtimrpt.sql report lists all workers for AD Parallel jobs separately. So the top 100 can be dominated by a handful of jobs. You can use the following SQL to list all jobs in order of maximum elapsed time (descending), but reporting all workers of an AD Parallel job in one line. Note that <session_id> is the ID for the upgrade session and not a user session.
60
Also beware of jobs that are called multiple times in the same phase (e.g. akload.class, LoadMap.class, XDOLoader.class). Finally, note that the script only reports completed jobs.
SELECT phase, phase_name, product, job_name, max_elapsed_time, min_start_time, max_end_time, workers FROM (SELECT phase, phase_name, product, job_name, MAX(elapsed_time) elapsed_time_unconv, LPAD(FLOOR(MAX(elapsed_time)*24), 4)||':'|| LPAD(FLOOR((MAX(elapsed_time)*24-floor(MAX(elapsed_time)*24))*60), 2, '0')||':'|| LPAD(MOD(ROUND(MAX(elapsed_time)*86400), 60), 2, '0') max_elapsed_time, INITCAP(TO_CHAR(MIN(start_time),' MON DD HH24:MI', 'NLS_DATE_LANGUAGE = american')) min_start_time, INITCAP(TO_CHAR(MAX(end_time),' MON DD HH24:MI', 'NLS_DATE_LANGUAGE = american')) max_end_time, count(worker_id) workers FROM ad_task_timing WHERE session_id = <session_id> GROUP BY phase, phase_name, product, job_name) ORDER BY elapsed_time_unconv DESC;
Matching long-running SQL with a job
The following variant will give jobs running at any point between two time intervals, with the longest running jobs first. This is useful in matching SQLs to jobs: <start_time> = start of period to report in format YYYYMMDDHH24MISS <end_time> = end of period to report in format YYYYMMDDHH24MISS Note that the job must have completed for it to be reported by this script.
SELECT phase, phase_name, product, job_name, max_elapsed_time, min_start_time,
61
max_end_time, workers FROM (SELECT phase, phase_name, product, job_name, MAX(elapsed_time) elapsed_time_unconv, LPAD(FLOOR(MAX(elapsed_time)*24), 4)||':'|| LPAD(FLOOR((MAX(elapsed_time)*24-floor(MAX(elapsed_time)*24))*60), 2, '0')||':'|| LPAD(MOD(ROUND(MAX(elapsed_time)*86400), 60), 2, '0') max_elapsed_time, INITCAP(TO_CHAR(MIN(start_time),' MON DD HH24:MI', 'NLS_DATE_LANGUAGE = american')) min_start_time, INITCAP(TO_CHAR(MAX(end_time),' MON DD HH24:MI', 'NLS_DATE_LANGUAGE = american')) max_end_time, count(worker_id) workers FROM ad_task_timing WHERE session_id = <session_id> AND ( start_time BETWEEN TO_DATE('<start_time>','YYYYMMDDHH24MISS') AND TO_DATE('<end_time>','YYYYMMDDHH24MISS') OR NVL(end_time, start_time+elapsed_time) BETWEEN TO_DATE('<start_time>','YYYYMMDDHH24MISS') AND TO_DATE('<end_time>','YYYYMMDDHH24MISS') ) GROUP BY phase, phase_name, product, job_name) ORDER BY elapsed_time_unconv DESC;
To find upgrade AD jobs that are in progress, you can use adctrl option 1 (Show worker status). You can tell when they started by looking at the patch log file. e.g.
$ cat u_merged.log|grep -A2 cstpostimportaad.sql Assigned: file cstpostimportaad.sql on worker 48 for product bom username BOM. Time is: Fri Mar 22 2013 22:48:54
Report of Jobs Running at Different Times During The Upgrade
You often want to know: Which jobs were running at a particular time: this helps identify the job that a long running SQL (in AWR) belongs to. The particular jobs causing events (such as contention/waits) on AWR.
62
Which jobs were running at the same time as another job. Which jobs were running during a long period of low worker utilization. However, be aware that this may be legitimate: for example, a job that runs parallel SQL/DML, or creates objects in parallel.
The following script reports on the jobs that were running at specified intervals during the upgrade. < session_id> = Id for the upgrade session <interval> = The interval in minutes <start_time> = start of period to report in format YYYYMMDDHH24MISS <end_time> = end of period to report in format YYYYMMDDHH24MISS We use this date/time format because it is same in all regions and has no delimiters. -1 for <start time> and <end time> reports the whole period of upgrade session -1 for <end time> just reports the jobs running at the start time Note that this script does not show all jobs run in any interval - just those that were running at the sample times. We use FND_TABLES as a dummy table to generate interval rows - this will give a maximum of around 20,000 intervals, so you should choose an interval size that does not result in more intervals than this.
BREAK ON time_slot WITH ad_start_end AS (SELECT MIN(start_time) start_time, MAX(end_time) end_time FROM ad_task_timing WHERE session_id = <session_id> ) ,intervals AS (SELECT (NVL(TO_DATE(DECODE(<start_time>,'1',NULL,<start_time>),'YYYYMMDDHH24MISS'),st.start_time) + (((rownum-1)* TO_NUMBER(<interval>))/(24*60))) interval_time FROM fnd_tables ft, ad_start_end st WHERE rownum <= DECODE(<start_time> ,'-1' ,CEIL(((st.end_timest.start_time)*24*60)/TO_NUMBER(<interval>))+1 ,DECODE(<end_time>
63
,'-1' ,1 ,CEIL(((TO_DATE(DECODE(<end_time>,'1',NULL,<end_time>),'YYYYMMDDHH24MISS') -TO_DATE(DECODE(<start_time>,'1',NULL,<start_time>),'YYYYMMDDHH24MISS')) *24*60) /TO_NUMBER(<interval>))+1 ) ) ) SELECT TO_CHAR(di.interval_time,'DD-MON-YYYY HH24:MI:SS') time_slot, adt.job_name job_running, COUNT(adt.worker_id) workers FROM intervals di, ad_task_timing adt WHERE di.interval_time BETWEEN adt.start_time AND NVL(adt.end_time, DECODE(adt.elapsed_time ,NULL ,di.interval_time ,adt.start_time+adt.elapsed_time) ) AND adt.session_id = <session_id> GROUP BY di.interval_time, job_name ORDER BY di.interval_time, job_name;
Report of worker utilization
The following script can be used to report worker utilization at specific intervals during the upgrade: <session_id> = ID for the upgrade session <interval> = The interval in minutes As with the previous script, we use FND_TABLES as a dummy table to generate interval rows - this will give a maximum of around 20,000 intervals, so you should choose an interval size that does not result in more intervals than this. You should be looking for long periods of time where worker utilization is much lower than the number of AD workers. However, be aware that it may be legitimate: for example, a job that runs parallel SQL/DML, or creates objects in parallel.
WITH ad_start_end AS (SELECT MIN(start_time) start_time, MAX(end_time) end_time FROM ad_task_timing WHERE session_id = <session_id> ) ,intervals AS
64
(SELECT (st.start_time + (((rownum-1)* TO_NUMBER(<interval>))/(24*60))) interval_time FROM fnd_tables ft, ad_start_end st WHERE rownum <= CEIL(((st.end_time-st.start_time)*24*60)/TO_NUMBER(<interval>))+1 ) SELECT TO_CHAR(di.interval_time,'DD-MON-YYYY HH24:MI:SS') time_slot, COUNT(adt.worker_id) workers, DECODE(MIN(adt.job_name),MAX(adt.job_name),MIN(adt.job_name),'Multiple') job_running FROM intervals di, ad_task_timing adt WHERE di.interval_time BETWEEN adt.start_time AND NVL(adt.end_time, DECODE(adt.elapsed_time ,NULL ,di.interval_time ,adt.start_time+adt.elapsed_time) ) AND adt.session_id = <session_id> GROUP BY di.interval_time ORDER BY di.interval_time;
Report of batch/worker utilization for a specific job
The following script can be used to report the number of batches in progress for intervals during a specific job. It also shows if the size of batches or speed of processing is increasing or decreasing. <script_name> = script/job to be analyzed <interval> = the interval in minutes Again, we use FND_TABLES as a dummy table to generate interval rows - this will give a maximum of around 20,000 intervals, so you should choose an interval size that does not result in more intervals than this. Note that there could be previous upgrades with the same script name. So you may need to edit this script to ensure you only pick up data for the most recent run of the script. Also be aware that the script_name may not be the same as the job_name from the task timing reports. For example, it may contain an appended version number. The script name is usually assigned to the variable l_script_name in the actual script.
WITH ad_start_end AS (SELECT MIN(start_date) start_time, MAX(end_date) end_time FROM ad_parallel_update_units aduu, ad_parallel_updates adu WHERE aduu.update_id = adu.update_id AND adu.script_name = '<script_name>'
65
) ,intervals AS (SELECT (st.start_time + (((rownum-1)* TO_NUMBER(<interval>))/(24*60))) interval_time FROM fnd_tables ft, ad_start_end st WHERE rownum <= CEIL(((st.end_time-st.start_time)*24*60)/TO_NUMBER(<interval>))+1 ) SELECT adu.script_name, TO_CHAR(di.interval_time,'DD-MON-YYYY HH24:MI:SS') time_slot, COUNT(*) batches, ROUND(AVG(aduu.rows_processed),0) avg_row_process, ROUND(AVG((aduu.end_date - aduu.start_date)*(24*60*60)),0) avg_bat_tim_secs, DECODE(SUM(aduu.rows_processed),0,0,ROUND(((SUM(aduu.end_dateaduu.start_date))*(24*60*60)*10000)/(SUM(aduu.rows_processed)),0)) sec_per_10k_rows, ROUND(AVG(aduu.end_block + 1 - aduu.start_block),1) avg_blocks_per_bat FROM intervals di, ad_parallel_update_units aduu, ad_parallel_updates adu WHERE aduu.update_id = adu.update_id AND adu.script_name = '<script_name>' AND di.interval_time BETWEEN aduu.start_date AND NVL(aduu.end_date,di.interval_time) GROUP BY adu.script_name, di.interval_time ORDER BY adu.script_name, di.interval_time;
Report on CBO Statistics for all Oracle E-Business Suite tables
You may want to report on the CBO statistics for Oracle E-Business Suite tables during the upgrade, particularly before adsstats.sql is run. The script adsstats.sql will populate the statistics correctly before the end of the upgrade. The fact that tables have incorrect statistics during the upgrade will not be visible. So you may not be able to see that a table had null, zero or inaccurate CBO statistics, and that this is the reason for an expensive execution plan.
SELECT owner, table_name, num_rows, TO_CHAR(last_analyzed,'DD-MON-YYYY HH24:MI:SS') last_analyzed FROM all_tables WHERE owner IN (SELECT upper(oracle_username) sname FROM fnd_oracle_userid WHERE oracle_id BETWEEN 900 AND 999 AND read_only_flag = 'U' UNION ALL SELECT DISTINCT upper(oracle_username) sname FROM fnd_oracle_userid a, fnd_product_installations b WHERE a.oracle_id = b.oracle_id ) ORDER BY owner, table_name;
66
Analyze contention (waits) and where they occur
The following shows the sql_ids on which a particular wait occurs between two snapshots in AWR.
<db_id> is the database ID, <inst_num> is the instance number, <wait name> is the name of the wait, and <begin snap> and <end snap> are the start and end snapshot IDs.
select ss.sql_id, ss.time_waited, ss.counts_waited, tt.total_time, ROUND((ss.time_waited*100/tt.total_time),1) percent from (select s.sql_id , COUNT(*) counts_waited, SUM(time_waited) time_waited from DBA_HIST_ACTIVE_SESS_HISTORY s, DBA_HIST_SEG_STAT_OBJ o where s.dbid = <db_id> and s.instance_number = <inst_num> and o.dbid (+) = s.dbid and o.obj# (+) = s.current_obj# and s.event = '<wait name>' and snap_id between <begin_snap> and <end_snap> group by s.sql_id) ss ,(select SUM(time_waited) total_time from DBA_HIST_ACTIVE_SESS_HISTORY t where t.dbid = <db_id> and t.instance_number = <inst_num> and t.event = '<wait name>' and t.snap_id between <begin_snap> and <end_snap>) tt order by ss.counts_waited desc;
The following SQL shows the objects on which a particular wait occurs for a given SQL ID (between two snapshots in AWR). Where <db_id> is the database ID, <inst_num> is the instance number, <wait name> is the name of the wait, <sql_id> is the SQL ID and <begin snap>, and <end snap> are the start and end snapshot IDs.
select ss.sql_id, ss.event, ss.current_obj#, ss.owner, ss.object_name, ss.object_type, ss.time_waited, ss.counts_waited, tt.total_time, ROUND((ss.time_waited*100/tt.total_time),1) percent from (select s.sql_id, s.event, s.current_obj#, o.owner, o.object_name, o.object_type , COUNT(*) counts_waited, SUM(time_waited) time_waited from DBA_HIST_ACTIVE_SESS_HISTORY s, DBA_HIST_SEG_STAT_OBJ o where s.dbid = <db_id> and s.instance_number = <inst_num> and s.sql_id = '<sql_id>' and s.event = '<wait name>' and o.dbid (+) = s.dbid and o.obj# (+) = s.current_obj# and snap_id between <begin_snap> and <end_snap> group by s.sql_id, s.event, s.current_obj#, o.owner, o.object_name, o.object_type) ss ,(select SUM(time_waited) total_time from DBA_HIST_ACTIVE_SESS_HISTORY t
67
where t.dbid = <db_id> and t.instance_number = <inst_num> and t.sql_id = '<sql_id>' and t.event = '<wait name>' and t.snap_id between <begin_snap> and <end_snap>) tt order by ss.counts_waited desc
Common Solutions
Known Issues
Once you have identified the long running jobs and SQL, you should check My Oracle Support for known issues and potential solutions or workarounds. However, bear in mind that a fix or workaround may not necessarily fix your particular problem. Some bugs and SRs do not show the symptoms of the performance problem, but only the solution. If possible, you should obtain evidence (diagnostics) to justify the suggested fix. If you cannot confirm that you have exactly the same issue (from the diagnostics). You may still apply the fix, but should continue to gather diagnostics and search for a solution until the issue is fully resolved. We have encountered many situations where a customer has focused on a particular known issue and fix, only to find out that their performance issue was different from the known one and therefore the required fix was different. This is because each customer uses the Oracle E-Business Suite modules and functionality in a different way, and has different volumes and distributions of data. They may also have a different database version or configuration, hardware configuration, and CBO and database initialization parameters.
Custom Indexes
There will be cases where a long-running job has an inefficient execution plan, and so a new index will have an enormous impact. Typically, the execution plan will use a full table scan, an unselective index range scan, or an index skip scan, all of which can result in a large number of rows being filtered out when the table is accessed. In such cases, you could create a custom index to filter out the rows, reducing the number of rows accessed on the table or the need for an index skip scan.
68
Ensure you still raise an SR for the performance issue (supplying diagnostic evidence). The performance issue may still need to be resolved in the standard code.
SQL Profiles for Inefficient Execution Plans
If, when using the diagnostics above (especially display cursor report or SQL Trace), you identify that a long running job has an inefficient execution plan, you could use a SQL Profile to apply hints that will help the CBO choose a better execution plan. You will need SQL tuning expertise to do this. Ensure you still raise an SR for the performance issue (with diagnostic evidence). The performance issue may still need to be resolved in the standard code. The syntax for hints in SQL Profiles is stricter than for hints applied directly in the SQL. If you get them wrong, they are just ignored. You can still use lower and upper case, omit quotes (), and leave out the leading query block names. However, the best approach is to: 1. 2. Run EXPLAIN PLAN on the SQL (with hints added). Get the fully formed hints from the outline by running:
SELECT * FROM TABLE(dbms_xplan.display(FORMAT => 'ALL +OUTLINE')).
This helps get the correct syntax. 3. 4. Create the SQL profile script and execute it. Run EXPLAIN PLAN on the SQL (without hints added). This allows you to see if the SQL Profile has been applied (the Notes section at the end of the output will contain SQL profile "<profile_name> used for this statement), and the explain plan should show that the hints have been applied.
Here is an example of a script to create a SQL Profile, applying hints.

DECLARE l_sql_fulltext clob := NULL; lv_hint SYS.SQLPROF_ATTR := SYS.SQLPROF_ATTR (); BEGIN l_sql_fulltext := 'SELECT AAGC.APPLICATION_ID, AAGC.APPROVAL_GROUP_ID FROM AME_APPROVAL_GROUP_CONFIG AAGC WHERE SYSDATE BETWEEN AAGC.START_DATE AND NVL(AAGC.END_DATE - (1/86400), SYSDATE) AND NOT EXISTS (SELECT NULL FROM AME_ACTION_USAGES AAU, AME_RULE_USAGES ARU, AME_ACTIONS AA, AME_ACTION_TYPES AAT WHERE AA.ACTION_ID = AAU.ACTION_ID AND AAU.RULE_ID = ARU.RULE_ID AND ARU.ITEM_ID =
69
AAGC.APPLICATION_ID AND SYSDATE BETWEEN AA.START_DATE AND NVL(AA.END_DATE - (1/86400), SYSDATE) AND SYSDATE BETWEEN AAT.START_DATE AND NVL(AAT.END_DATE - (1/86400), SYSDATE) AND AA.PARAMETER = TO_CHAR(AAGC.APPROVAL_GROUP_ID) AND AAT.ACTION_TYPE_ID = AA.ACTION_TYPE_ID AND AAT.NAME IN (''pre-chain-of-authority approvals'' , ''post-chainof-authority approvals'' , ''approval-group chain of authority'') AND ROWNUM < 2) ORDER BY AAGC.APPLICATION_ID'; lv_hint.EXTEND; lv_hint(1) := 'BEGIN_OUTLINE_DATA'; lv_hint.EXTEND; lv_hint(2) := 'USE_NL(@"SEL$2" "AAT"@"SEL$2")'; lv_hint.EXTEND; lv_hint(3) := 'USE_NL(@"SEL$2" "ARU"@"SEL$2")'; lv_hint.EXTEND; lv_hint(4) := 'USE_NL(@"SEL$2" "AAU"@"SEL$2")'; lv_hint.EXTEND; lv_hint(5) := 'LEADING(@"SEL$2" "AA"@"SEL$2" "AAU"@"SEL$2" "ARU"@"SEL$2" "AAT"@"SEL$2")'; lv_hint.EXTEND; lv_hint(6) := 'INDEX(@"SEL$2" "AAT"@"SEL$2" ("AME_ACTION_TYPES"."ACTION_TYPE_ID" "AME_ACTION_TYPES"."START_DATE" "AME_ACTION_TYPES"."END_DATE"))'; lv_hint.EXTEND; lv_hint(7) := 'INDEX(@"SEL$2" "ARU"@"SEL$2" ("AME_RULE_USAGES"."RULE_ID" "AME_RULE_USAGES"."START_DATE" "AME_RULE_USAGES"."END_DATE"))'; lv_hint.EXTEND; lv_hint(8) := 'INDEX(@"SEL$2" "AAU"@"SEL$2" ("AME_ACTION_USAGES"."ACTION_ID"))'; lv_hint.EXTEND; lv_hint(9) := 'INDEX(@"SEL$2" "AA"@"SEL$2" ("AME_ACTIONS"."PARAMETER"))'; lv_hint.EXTEND; lv_hint(10) := 'FULL(@"SEL$1" "AAGC"@"SEL$1")'; lv_hint.EXTEND; lv_hint(11) := 'LEADING(@"SEL$1" "AAGC"@"SEL$1")'; lv_hint.EXTEND; lv_hint(12) := 'IGNORE_OPTIM_EMBEDDED_HINTS'; lv_hint.EXTEND; lv_hint(13) := 'END_OUTLINE_DATA'; dbms_sqltune.drop_sql_profile (name => 'R1211_AMEMIGCFG_1' ,ignore => TRUE ); dbms_sqltune.import_sql_profile( sql_text => l_sql_fulltext ,category => 'DEFAULT' ,name => 'R1211_AMEMIGCFG_1' ,profile => lv_hint ,description => 'R1211 amemigcfg.sql ,force_match => TRUE ); END; /
23y6y8d0r9v61'
70
You could also use SQL Baselines to import an execution plan from cursor cache, AWR or SQL Tuning set. However, this does not have the ability to import hints. Note that the SQL Tuning Advisor typically uses hints to apply statistical fixes (using OPT_ESTIMATE and others), rather than specifying the execution method. In conclusion, SQL Profiles:
Are easy to apply. Can be used to directly specify the join order, access and join methods. Are very stable.
Pragmatic Stable Execution Plans
In most cases, there is one join order that will give a good execution plan and minimize throwaway (that is, unnecessary access to rows that are filtered out at a later stage). Often, that join order is the natural one, following the flow of the application.
AD Parallel Jobs
AD Parallel jobs will access batches of rowids from a driving table (specified when calling AD_PARALLEL_UPDATES_PKG.INITIALIZE_ROWID_RANGE to initialize AD Parallel). In almost all cases these batches will contain a very small percentage of the total rows. So the execution plan should lead with the driving table (accessing it by rowid), and then use nested loop join method and index access. This will give a pragmatic plan for AD Parallel jobs. It may not be the very best execution plan, but it will not be particularly inefficient. For small tables (particularly lookup or reference), a full table scan and hash join may be better, but (compared to the overall workload) the difference could be negligible. You will typically use rowid, leading, use_nl and index hints. You could also use the undocumented cardinality hint to specify a low cardinality for the driving table (e.g. cardinality(ai 1).
Parallel SQL
For jobs that use Parallel SQL, full table scans and hash joins will usually but not always, - be better. They will also benefit from using a join order that minimizes throwaway.
71
Long Running Index Creation
This will typically apply to xdf and odfs. You should obtain an AWR report for the snapshots where the index is being created.
Not running with parallel DML
First of all check that the index is being created in parallel (e.g. CREATE INDEX .. PARALLEL) when the odf file is processed by AutoPatch. The DDL command will show up in the AWR (SQL Ordered By). If the index is being created in serial then you may need to increase the AutoPatch parameter parallel_index_threshold. This parameter specifies the number of blocks in a table. If a table contains fewer blocks than the threshold setting, indexes are created with parallel AD workers and serial DML (each worker creating a different index). If the table contains more blocks than the threshold setting, indexes are created with one worker and parallel DML. The valid values are 0 to 2147483647. If set to 0, indexes are created with parallel workers and serial DML. Default value: 20000; meaning a threshold of 20,000 blocks. Otherwise you may need to gather the statistics for the table. If the statistics are missing or incorrect (e.g. too low) then AutoPatch (adodfcmp) may get the wrong number of blocks and choose to create with serial DML. If you cannot get the index to create in parallel, you should consider pre-creating the index see below.
parallel_max_servers / parallel degree
Check if parallel_max_servers needs increasing. If the SQL is not using as many parallel slaves as expected, then it is possible that the number of parallel slaves is being limited by other simultaneous tasks. You can check AWR for any simultaneous tasks taking place at the same time as the create index. Another limitation could be the CPU_COUNT * PARALLEL_THREADS_PER_CPU parameters. CPU_COUNT, if set, should either be the number of CPU cores or zero. Note that if CPU_COUNT is not set, or set to zero, it will default to the number of CPUs reported by the operating system. This is therefore acceptable.
72
PARALLEL_THREADS_PER_CPU if present should be 2. If it is not set, the default is platform-dependent (normally 2), and adequate in most cases. Note that Automatic Parallel Degree Policy (PARALLEL_DEGREE_POLICY) should not be enabled - the default is MANUAL. See Oracle Database VLDB and Partitioning Guide 11g Release 2 (11.2). Part Number E25523-01, How Parallel Execution Works and Tuning General Parameters for Parallel Execution sections.
pga_aggregate_target
Create index will use space in the PGA for sorting. If (and only if) there is significant usage of temporary space should you consider increasing pga_aggregate_target.
Contention between parallel slaves
If the index is being created in parallel (and with enough slaves), the performance issue may be caused either by contention between the parallel slaves creating the index or contention with another simultaneous process. Note that if the parallel degree is too high this will also cause contention. Obtain an AWR report for the period when the index is being created and check for the contention (waits, CPU usage). If there is no contention on AWR, it is unlikely that pre-creating the index will resolve the issue. You would be using the same command to pre-create, with the same parallel degree.
Pre-Creating
If the index is being created (by odf) in serial, and either you cannot get it to create in parallel or there is significant contention when the index is created (by odf) in parallel, then pre-creating the index is likely to help. However, there are some points to bear in mind: Do not do it too soon. If the table subsequently has heavy DML (insert, delete or update of a column used in the index) prior to where the normal creation would have occurred (with odf), then that DML is likely to take much longer. So if possible, pre-create the index after any job that executes heavy DML on the table.
73
Also, be careful to create the index with exactly the same definition as in the odf file. adodfcmp will compare the index that has already been pre-created with the definition in the odf file. If they do not match, adodfcmp may need to alter or re-create the index anyway. Remember to ALTER INDEX to revert the degree of parallel (e.g. NOPARALLEL) afterwards.
So if you do want to pre-create indexes manually, you should do so as late as possible.

High level of contention on indexes on long running DML jobs
If a long running job running heavy DML (typically INSERT, but it could be UPDATE or DELETE), has a high level of contention, and that contention is on particular indexes, you could consider dropping the indexes before the job and re-creating them afterwards (provided that the index is not used by any execution plan in the meantime). Alternatively, you could consider dropping the indexes prior to the upgrade and re-creating them afterwards. However, you need to be sure that the indexes are not being used during the upgrade. You can check if indexes have been used during any particular period by using the command 'ALTER INDEX <index> MONITORING USAGE', and then querying the view V$OBJECT_USAGE. In addition, the AWR table DBA_HIST_SQL_PLAN will tell you what indexes are in use, and DBA_HIST_ACTIVE_SESS_HISTORY tells you how much time is used in maintaining the indexes for each piece of SQL. You should ensure that you re-create in parallel and with exactly the same definition. And remember to ALTER INDEX to revert the degree of parallel (e.g. NOPARALLEL) afterwards. When deciding this, you need to take into account the percentage of rows being inserted/deleted/updated during the DML, as re-creating the index will be for all rows. This will be particularly useful on custom indexes, where we advise dropping the indexes prior to the upgrade and re-creating them afterwards (provided that the custom index does not make any of the upgrade jobs significantly quicker).
High Level of enq: HW contention or enq: HV contention
This is related to the case above. If a long-running job inserting into a table and indexes has a high level of waits (enq: HW contention or enq: HV contention) you could do the following:
74
If the wait is occurring largely on particular indexes, drop the indexes before the job and recreate them afterwards (as above), provided the index is not used in the meantime. Increase the extent size on the table and indexes. Pre-allocate extents for the table and indexes. Partition the table and any indexes to spread the load across multiple partitions. Note that the tables and indexes will still be partitioned after go live. So only do this if the partitioning method will also give benefits after going -live on the production environment).
These enq: HW contention or enq: HV contention are often accompanied by higher levels of other waits such as enq: TX contention or buffer busy waits. The HW enqueue is used to manage the allocation of space beyond the high water mark of a segment. So the wait enq: HW contention typically means that sessions are waiting for the allocation of extents. The wait enq: HV contention is the same, except it occurs when parallel slaves are waiting for an extent to be allocated. You can query the AWR tables directly to get more detail on where waits are occurring (e.g. on which object). See the Useful Scripts section.
High Level of redo log waits log buffer space, log file sync etc.
If you have a high level of waits associated with redo log, especially log buffer space and log file sync, you could consider changing the configuration of redo logs; moving to a faster filer; increasing the size or number of logs; or increasing the log parallelism (hidden initialization parameter _log_parallelism_max). You could consider running with NOLOGGING, but this is not advised. It would require turning logging off for all tables and indexes and then switching it back on post-upgrade for all the tables and indexes where it had been switched off.
Long Running Statistics Gathering (adsstats.sql)
If the upgrade is taking a long time to gather CBO statistics (adsstats.sql), you could consider the following strategies: Increasing parallel_max_servers and pga_aggregate_target (adsstats.sql should run with few or no concurrent tasks).
75
Using all nodes on Oracle RAC system. Importing statistics gathered during test runs.
Note that the last suggestion is a strategic action that you may take for all Release 12 upgrades. So it is considered and described in more detail in the Preparation and Planning section.
Gathering or Deleting Statistics to resolve specific performance issues
Performance issues may indicate the need for you to gather or delete statistics for specific objects (see Resolving Performance Issues section).
New Tables
The Release 12 upgrade creates and populates many tables. For most tables the statistics are not gathered until late in the upgrade (i.e. by adsstats.sql in the last+63 phase of the R12.1.1 upgrade). So the statistics may be empty (for new Release 12 tables), and dynamic sampling may not have resulted in a good execution plan. Alternatively, the statistics may be incorrect. In such a case, you may want to gather statistics (using FND_STATS.GATHER_TABLE_STATS) for a specific table at a particular stage of the upgrade. However, you should only gather statistics for specific tables when you know the following: Incorrect or missing statistics on that table is definitely causing the performance issue. There are no significant inserts to the table afterwards (or significant changes in the number of distinct values (NDV) on key join/filter columns or the distribution of values on columns with a histogram). In such cases, the CBO statistics will then be incorrect and could result in subsequent poor execution plans.
Remember that having no statistics can be better than incorrect statistics (as dynamic sampling will be used) and they are certainly better than zero statistics (gathered when table was not populated). You can also seek guidance about gathering statistics from Oracle Support.
Temporary Tables
If performance issues are due to transient or temporary tables (including Global Temporary tables), and the statistics for these tables contain unrepresentative statistics, you may want to do one of two things:
76
Before the upgrade, delete the statistics using DBMS_STATS.DELETE_TABLE_STATS, lock using DBMS_STATS.LOCK_TABLE_STATS, and then exclude using FND_STATS.LOAD_XCLUD_TAB. You can then rely on dynamic sampling. Gather representative statistics using FND_STATS.GATHER_TABLE_STATS at an appropriate point or import representative statistics using FND_STATS.RESTORE_TABLE_STATS (with CASCADE = TRUE/default). Then lock using DBMS_STATS.LOCK_TABLE_STATS and exclude using FND_STATS.LOAD_XCLUD_TAB. The representative statistics will have been previously exported using FND_STATS.BACKUP_TABLE_STATS (with CASCADE = TRUE/default).
See the Preparation and Planning - Pre-Upgrade Environment/Activities, Gather CBO Statistics, Zero or Incorrect Statistics on transient /temporary tables section.
Long-Running Materialized View xdf/odfs
If you have long-running xdf or odf jobs creating materialized views (MV), consider cleaning up or truncating of any large MV logs (note that this requires MV complete refresh).
Long-Running Jobs that might not be needed
Check if any long-running jobs are actually needed on your system, especially if they do not alter any data (insert, update or delete any rows).It could be that some jobs are for a module or localization that you do not use, or are not even licensed for. Oracle Support may advise that you can skip a job, or offer a fix or workaround for you case.
Skipping Jobs
If Oracle Support has identified that a particular job is not required for your instance, there are two options for skipping it: You can comment the job out of the unified driver (u driver) for the install. You can run the adctrl utility and choose hidden option 8.
Maximum AD Parallel Workers
You should be aware that AutoPatch specifies a maximum number of workers. The number depending on several factors, and under certain circumstances may be less than the number of workers you require. So you will need to resolve the issue.
77
You will get an error message similar to the following: AD utilities can support a maximum of 999 workers. Your current database configuration supports a maximum of Z workers. Oracle recommends that you use between X and Y workers. In the above message, X, Y and Z are calculated as follows: X = 2 * cpu_count Y = 4 * cpu_count Z = (processes - ((total rows of v$process) + parallel_max_servers)) / 2 Where: processes = processes db initialization parameter. parallel_max_servers = parallel_max_servers db initialization parameter. total rows of v$process = select count(1) from v$process cpu_count = cpu_count db initialization parameter (or default value). So, the maximum is basically 50% of the available "processes" (from the database initialization file), once the current number of processes (rows in v$processes) and the maximum that could be used by Parallel Query/DML have been deducted. See My Oracle Support Knowledge Document 800024.1 "How Does Adpatch Determine The Number Of Workers To Recommend?"
78
Best Practices for Minimizing Oracle EBusiness Suite Release 12 Upgrade Downtime Jun 2013 Author: Jim Machin Editor: Robert Farrington 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
Copyright 2013, 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 and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners. Intel and Intel Xeon are trademarks or registered trademarks of Intel Corporation. All SPARC trademarks are used under license and are trademarks or registered trademarks of SPARC International, Inc. AMD, Opteron, the AMD logo, and the AMD Opteron logo are trademarks or registered trademarks of Advanced Micro Devices. UNIX is a registered trademark of The Open Group. 0113

Best Practices For Minimizing Oracle EBS R12 Upgrade Downtime - Final

Uploaded by

Copyright:

Available Formats

Best Practices For Minimizing Oracle EBS R12 Upgrade Downtime - Final

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Best Practices For Minimizing Oracle EBS R12 Upgrade Downtime - Final

Uploaded by

Copyright:

Available Formats

What are the main steps for preparing for an E-Business Suite upgrade?

What are the main steps for preparing for an E-Business Suite upgrade?

What are some common performance issues that can occur during an upgrade?

What are some common performance issues that can occur during an upgrade?

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

An Oracle White Paper June 2013

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Common Solutions ...................................................................... 68

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Both are these are discussed below.

Preparation and Planning

Resolving Performance Issues

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Prediction of Performance Issues

Preparation and Planning - Pre-Upgrade Environment and Activities

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Preparation and Planning Release 12 Upgrade Activities

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

For more than 32 cores, start with:

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Fixed Object Statistics should be gathered:

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

After the Oracle E-Business Suite Upgrade

Resolving Performance Issues - Using Diagnostics

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Resolving Performance Issues Common Solutions

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Pragmatic Stable Execution Plans

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Long Running Statistics Gathering (adsstats.sql)

Gathering or Deleting Statistics to resolve specific performance issues

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Key Resources and Documents

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

bde_last_analyzed.sql - Verifies CBO Statistics (Doc ID 163208.1)

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Preparation and Planning - Pre-Upgrade Environment / Activities

OATM (Oracle Applications Tablespace Model)

Multiple Organizations (Multi-Org) Architecture

Purge old data

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Purge interface data

Gather CBO Statistics

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Fixed Object and Dictionary Statistics

Drop MRC Schema

Preparation and Planning Release 12 Upgrade Activities

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

Other recommended performance fixes (if not already applied) are:

Removing Unnecessary Workloads / Overheads

Best Practices for Minimizing E-Business Suite R12 Upgrade Downtime

AND fsh.last_gather_end_time IS NOT NULL ORDER BY (fsh.last_gather_end_time - fsh.last_gather_start_time)243600 desc