Upgrading Oracle Database 9i to 10g for Siebel usin g SQL Performance Analyzer (SPA)
Sameer Marwa – Infogig ConsultingKhaled Yagoub – Oracle Development Ravi Sharma – Oracle Consulting
Outline
� The Beginning and Challenges� Proposed Solution� SQL Performance Analyzer (SPA) & Traditional Approach� Modified Test Approach� Strategy & Results: Production Run� Lessons learned & Future plans� Q/A
The Beginning
Business Requirements
� Upgrade SIEBEL database containing critical customer data, orders and service requests, from 9i to 10g
And by the Way
� Provide ability to failback to 9i with no data loss
� Upgrade to be completed within the maintenance window (6hrs)
� No screw up like the last time (last CBO refresh resulted in DB outages!)
Devil is in the Details
Challenges:
� Very large environment: 10TB sized database on dual node HP Superdome servers cluster (96 CPUs each), 25+ Application servers
� High transaction volume: 15k concurrent users
� Large number of SQL to analyze: 100k unique SQL (Initial estimates indicated that 50% of these would have different execution planswith new 10g CBO)
� Modifications/enhancements to the SIEBEL application often changes 20-30% of the SQL
� Minor application release was planned 4 weeks before the 10g Upgrade
Devil is in the Details
Challenges:
� Expensive downtime: $$$ per hour
� Lengthy Database/Application restart: Each restart requires 2 hours ($$$ per restart)
� Hosted environment: multiples teams added to the complexity
Get Jack out of the Box
Proposed Solution
� Build a new 10g environment while maintaining the 9i system, allowing a failback to 9i in case of a problem : The 10g DB maintained in sync with the 9i environment using GoldenGate
� Use SQL Performance Analyzer (SPA) to mitigate SQL related concerns (End to end analysis should take no more than 2 business days)
What is SPA?
� Utility that predict the impact of system changes on SQL workload response time
� SQL statements are captured and stored in SQL Tuning Set (STS)
� SQL performance data before and after change is generated using:
� Method 1: EXECUTE SQLs� Method 2: GENERATE PLANs� Method 3: CONVERT SQLSET
� Regressed SQL are identified by comparing the execution statistics(buffer gets, cpu time, etc)
Analysis Report
Compare
SQL Performance
SQL plans + stats SQL plans + stats
Pre-change SQL Trial
Post-change SQL Trial
SQL Tuning Set (SQL Workload: SQL text + binds)
What is SPA?
� Replays individual SQL but not concurrent SQL workload
� Common System change scenarios� Database upgrades and patches
� Database parameter changes
� Schema changes� Statistics gathering
� Implementation of tuning recommendations
� OS/hardware changes
� Integrated with SQL tuning set, SQL plan baselines, SQL tuning advisor, and Enterprise Manager to provide an End-to-end solution
SPA: Traditional Approach for 9i to 10g Upgrade*
1. Use SQL_TRACE to capture SQL statements on 9i production system � Trace database sessions to capture application related SQL with related bind
variables and execution statistics
2. Build SQL Tuning Set (STS) from SQL_TRACE files
3. Build SQL Trial Baseline: 9i SQL plans and statistics � Convert SQL tuning set into a SPA SQL Trial
4. Build Post-upgrade SQL Trial : 10g SQL plans and statistics � Using EXECUTE method to remotely execute SQL against the 10g database
5. Compare baseline to post-upgrade SQL trials� Identify SQL STATEMENTS that regressed (modifiable thresholds) in 10g
6. Tune the regressed SQL statements � Using SQL profiles, Stored Outlines, rewrite the SQL, change the CBO statistics
7. Verify fixes by re-running SPA and comparing baseline to post-tuning SQL trials
*OTN: Testing Performance Impact of an Oracle Database 9i/10g Release 1 to Oracle Database 10g Release 2 Upgrade with SPA
Challenges With the Traditional Approach
� Potential impact of SQL_TRACE to transaction response time
� Very difficult determine relevant sessions to trace
� Potential high number of sessions to trace and capture a representative set of application SQL statements
� Potential impact (CPU and disk space for trace files) of tracing difficult to quantify
� Expect a very large number of SQL statements to test with
SPA Approach with Few Modifications
1. Use SQL_TRACE to capture SQL statements on 9i production system
2. Build SQL Tuning Set (STS) from SQL_TRACE files
3. Build SQL Trial baseline: 9i SQL plans and statistics
4. Build post-upgrade SQL Trial: 10g SQL plans and statistics using EXECUTE Method
5. Compare baseline to post-upgrade SQL trials to identify regressed SQL
6. Tune the regressed SQLs
7. Verify fixes
1. Manually query v$sql_xxx views to capture SQLs and plans on 9i production system
2. Manually build STS from v$sql_xxx resutls
3. Convert STS to build SQL Trial Baseline: 9i
4. Build post-upgrade SQL Trial using GENERATE PLAN method: plans only
5. Compare 9i and 10g plans to identify Changed Plans
6. Manually analyze changed plans to identify critical plan differences
7. Use SQL_TRACE on 9i prod. to capture BINDS for only the critical SQL subset
8. Use SPA EXECUTE method for the critical subset of SQL to identify regressions
9. Compare …… etc.
SPA Traditional Approach SPA Modified Approach
SPA Approach with Few Modifications: Setup
10g Test / Production DB Cluster
9i Production DB Cluster
SPA Server 11g
(4 CPU win2k)
SPA ProcessManual Process� Remotely Get Exec. Plans for
All SQLs using DB Link
� Compare with 9i Plans
� Execute SQL with Bind Values
� Find Regressed SQLs
� Get Exec. Plans for All SQLs
� Get Bind Values for Critical SQLs
Production Run: Details
Step 1:Captured SQLs with related Plans and execution statistics
� Query v$sql, v$sqltext_withnewlines, and v$sql_plans
� Select the following statement types: SELECT, UPDATE and DELETE
� Ignore INSERT since then do not have sub-queries and hence will have same performance on 10g as on 9i
� Create two tables to save the results:• STATEMENTS: to store SQL text (CLOB),
parsing schema name, number of executions, elapsed time, buffer gets, etc.
• PLANS: to store SQL plan lines
� Export the resulting two tables to the 11g DB used to run SPA
V$SQL V$SQLTEXT_WITHNEWLINES
V$SQL_PLANS
9i Production Database
Table PLANSTable STATEMENTS
11g SPA Server
Query
Export
Production Run: Details
Step 2:Built SQL Tuning Set (STS) from 9i SQL and related plans in the 11g SPA database
� Import tables PLANS and STATEMENTS into SPA Server
� Create an new empty SQL tuning set
� Run a script to manually populate the SQL tuning set from tables PLANS and STATEMENTS
11g SPA Server
SQL Tuning Set in 11g
115K Unique SQL
Table PLANSTable STATEMENTS
Import
Query
Production Run: DetailsCode Snippet
declare
stscur dbms_sqltune.sqlset_cursor;
Begin
dbms_sqltune.create_sqlset(‘ ALL_SQL_STS');
open stscur for
select sqlset_row(
sql_text => STATEMENTS.sql_text,
parsing_schema_name => STATEMENTS.parsing_schema_na me,
executions=> STATEMENTS.executions,
elapsed_time=> STATEMENTS.elapsed_time,
priority => STATEMENTS.seq,
cpu_time=> STATEMENTS.cpu_time,
plan_hash_value => 4294967295,
optimizer_cost => (select cost from PLANS where sql _seq = STATEMENTS.seq and id = 0),
sql_plan => (select cast(collect( sql_plan_row_type( NULL,NULL,NULL,NULL,
operation, options, object_node,object_owner, objec t_name,
NULL,NULL,NULL,optimizer, search_columns, id, paren t_id, NULL,
position, cost,cardinality, bytes, other_tag,partit ion_start,
partition_stop, partition_id,distribution, cpu_cost ,io_cost,
temp_space,access_predicates,filter_predicates, nul l,null,null,
null)) as sql_plan_table_type)
from PLANS where sql_seq = STATEMENTS.seq))
from STATEMENTS;dbms_sqltune.load_sqlset('STATEMENTS', stscur);
end;
Production Run: Details
Step 3:Built SQL Trial Baseline: 9i SQL Text, plans and statistics
� Use Enterprise Manager (EM) or DBMS_SQLPA package to convert SQL tuning set into a SPA SQL Trial
Step 4:Built Post-upgrade SQL Trial: 10g SQL with only execution plans
� Specify DB Link to connect to remote 10g system
� Use SPA GENERATE PLAN method to collect execution plans for all SQLs in the STS
10g Database System
Input
SQL Tuning Set in 11g
SPA in 11g
Baseline Trial
3. Convert
Post-upgrade Trial
4. Build Trial
Remotely Generate Plans
DB Link
declare
tname varchar2(30);begin
tname := dbms_sqlpa.create_analysis_task(task_name => ’10gUpgrade', sqlset_name=> ‘ ALL_SQL_STS',description =>'identify changed plans for 10g Upgra de');
dbms_sqlpa.execute_analysis_task(task_name => '10gUp grade', execution_name=> ' 9i_plans ', execution_type => ' CONVERT SQLSET');
dbms_sqlpa.execute_analysis_task( task_name => '10gUpgrade' , execution_name=> '10g_plans ', execution_type => ' EXPLAIN PLAN ', execution_params=> dbms_advisor.arglist(' DATABASE_LINK',
’ 10G_TEST_DB'));
dbms_sqlpa.execute_analysis_task( task_name=> '10gU pgrade', execution_name=>'compare plans', execution_type=>' COMPARE', execution_params = dbms_advisor.arglist(' COMPARISON_METRIC',
'OPTIMIZER_COST'));
end;
Production Run: DetailsCode Snippet
Production Run: Details
Step 5:Compared baseline and post-upgrade SQLtrials and Analyze SPA report
� SPA compares 9i and 10g execution plans� SPA identifies SQL with plan changes
Step 6:Manually analyzed changed plans to identify critical plan differences
� Used custom SQL queries on SPA views to identify SQL statement with potential performance violation:
• SQL executions plans that have different JOIN driving tables in 10g
• SQL plan involving certain undesirable operations such as INDEX FAST FULL SCAN, FULL TABLE SCAN, etc.
Comparison and analysis report
9i BaselineSQL Trial
10g SQL Trial
SQL PlansSQL Plans
11g SPA ServerSPA DBA Views
DBA/USER_ADVISOR_PLANSDBA/USER_ADVISOR_OBJECTSDBA/USER_ADVISOR_FINDINGS
Reality Check: Findings From Plan ComparisonBad News
� Very Large Set of SQL Captured• 115k unique SQL captured from
the 9i production database
� Large Number of Changed Plans• Close to 40% (38k) of the SQL
had different plans on 10g
� Still Unmanageable ! • Close to 35% (9k) of the
changed-plans had driving step differences
� It was very difficult to capture BIND VALUES for the critical set
115K
38K
9K
0 30000 60000 90000 120000
SQL withDifferent JoinDriving Table
SQL with PlanChanges
ALL SQLStatements
Number of SQL Statements
Reality Check: Findings From Plan ComparisonGood News
� Of the 9K SQL with driving step difference, only 347 SQL were executed more than 70 times/day !
� Number of SQL with high executions (>70 times /day ) is 3.2K
� SQL statements with executions >70 times /day represents 99%of ALL SQL executions
� They represents 69% of ALL buffer gets
� Of the SQL with executions < 70/day, we had only 37 SQL withFULL SCANS on large objects (with no such operation in 9i)
SQL with # executions < 10/day=108.9k
10/day < #execs < 70/day= 2.9k
#execs > 70/day= 3.2k= 2.8%
99% of all SQL executions & 69% of all buffer gets
Strategy for Production Run
� Limit the amount of work while still covering majority of the SQL workload
� Focus on SQL with high executions (>70/day ) SQL (SELECT, UPDATE & DELETE only)
� To address concerns with SQL with low executions but undesirable operations, include 37 SQL with FULL SCANS on large objects (with no such operations in 9i)
� Capture BIND VALUES for most often executed SQL using SQL_TRACE
SQL with # executions < 10/day=108.9k
10/day < #execs < 70/day= 2.9k
#execs > 70/day= 3.2k= 2.8%
99% of all SQL executions & 69% of all buffer gets
Production Run: DetailsStep 7:
Used SQL_TRACE on 9i production to capture BINDS for only Highly executed SQL� Attempt to capture BINDS using
SQL_TRACE for the highly executed SQL statements
� Export the resulting trace files into SPA 11g server
� Use SPA to convert SQL trace files into a SQL tuning set
� All SQL statements in the STS have bind values
� Some of highly executed SQL statements might might not be in the STS because they were not captured by SQL_TRACE
� Create a separate STS for SQL without bind values
3.2K
1.8K
1.4K
0 1000 2000 3000 4000
SQL withBindsValues
SQLwithout
Bind Values
HighlyExecuted
SQLStatements
Number of SQL Statements
Production Run: Details
Step 8:� Used SPA in GENERATE PLAN method to remotely collect 10g execution
plans for SQL without BIND VALUES
� Used SPA in EXECUTE method to remotely collect 10g execution plans and statistics for SQL with BIND VALUES
Step 9:� SQL without BIND VALUES
• Used SPA plan comparison method to identify SQL Statement with different execution plans in 10g (labeled changed-plan set)
• Used custom SQL queries on SPA views to identify SQL statements with potential performance violation
• e.g., SQL executions plans that have different driving tables in 10g
� SQL with BIND VALUES• Used SPA to identify SQL with regressed performance in 10g using
buffer_gets as comparison metric
Production Run: DetailsCode Snippet for EXECUTE Method
declare
tname varchar2(30);begin
tname := dbms_sqlpa.create_analysis_task(task_name => ’10gUpgradeExecute', sqlset_name=> ‘ HIGH_EXEC_SQL_WITH_BINDS',description =>'identify sql regressions for 10g Upgr ade');
dbms_sqlpa.execute_analysis_task(task_name => tname, execution_name=> ' 9i_exec ', execution_type => ' CONVERT SQLSET');
dbms_sqlpa.execute_analysis_task( task_name => '10gUpgrade' , execution_name=> '10g_exec ', execution_type => ‘ TEST EXECUTE', execution_params=> dbms_advisor.arglist(' DATABASE_LINK',
’ 10G_TEST_DB'));
dbms_sqlpa.execute_analysis_task( task_name=> tname , execution_name=>'compare plans', execution_type=>' COMPARE', execution_params = dbms_advisor.arglist(' COMPARISON_METRIC',
‘BUFFER_GETS'));
end;
Production Run: Result Details
GENERATE PLAN Method
SQL with
BIND VALUES
SQLwithout
BIND VALUES
1.4 K SQL 1.8 K SQL
SPA 11g
10g Test DB
Changed Plans1.6k SQL
Driving Step Difference
159 SQL
EXECUTE Method
SQL with
BIND VALUES
SQLwithout
BIND VALUES
1.4 K SQL 1.8 K SQL
SPA 11g
10g Test DB
Regressed SQLs
96 SQL
Production Run: Final Results
Step 10:Manually IDENTIFIED and TUNEDthe FINAL subset of potentially problematic SQL statements
� 96 Regressed SQL identified using EXECUTE method
� 159 SQL with driving steps difference identified using GENERATE PLAN method
� 37 SQL with FULL SCANS on large objects (with no corresponding operation in 9i)
� Developed 36 Stored Outlines to tune regressed SQL
� Used SPA to validate tuning changes
� Deployed tuning changes in 10g database
96 Regressed
SQL
Manual Analysis
159 Driving Step changed SQL 37 FULL
SCAN SQL
36 Stored Outlines
Go-Live with 10g!
� Only 6 SQL STATEMENTS had to be tuned post go-live on 10g!
� These SQL STATEMENTS had same execution plans as in 9i but different plan during execution on 10g (due to bind peeking!)
� System stable for 4 weeks with no database issues encountered. That is, until the next Siebel release was introduced into production!)
Lessons Learned: SPA Advantage !
� Very convenient to test with a large number of SQL statements
� Remote SQL Plan generation of 100k SQL took approximately 6 hours
� Plan comparison of 100k plans under 30 minutes !
� Flexible architecture to perform what-if analysis
� SQL workload and SQL trials are stored in the database which make it easy to query, build, and keep history of all SQL performance tests
� Database views with details of SPA inputs and results makes it easy to perform custom analysis and generate custom reports
• DBA/USER_ADVISOR_OBJECTS• DBA/USER_ADVISOR_PLANS• DBA/USER_ADVISOR_FINDINGS
Lessons Learned: But Customize Testing Approach
� Find a process to capture SQL and executions statistics from theproduction database that fits your environment:
• For 9i use SQL_TRACE to capture bind values for critical set (subset) of SQL statements
• For 10g use SQL Tuning Sets with appropriate filters and captureexecution frequency
� Large number of SQL will have changed execution plans in 10g • Prioritize the plan changes (driving table differences, unexpected
operation in the SQL plan, etc) that is relevant for your environment
� Execution method analysis (executing SQL with bind data) is a better indicator of SQL performance post change
� Execution method can still produce a large list of SQL (with regressed performance) to analyze
• Use comparison metric judiciously and/or use custom filters
We Are on 10g, Now What?
� No more SQL_TRACE!
� Use SPA to analyze the impact of new indices to be deployed into10g production system
� Use SPA to analyze the impact of refreshed CBO Statistics:
• Gather CBO Statistics in the 10g test system
• Capture SQL, related plans and execution statistics from the 10gproduction database using SQL Tuning Sets (STS)
• Import the production STS into the 11g SPA DB
• Use SPA (EXECUTE method) against the new CBO Statistics to identify regressed SQL