1. Closing the Query Processing Loop in Oracle 11g Allison Lee, Mohamed Zait

2. Agenda Motivation Solution Overview Cardinality Feedback Adaptive Cursor Sharing General Feedback Process Experimental Results Conclusions

3. Motivation

4. Two query optimization problems: Estimating cardinality of intermediate results Optimizing queries containing query variables

5. Problem – Cardinality Estimation Cardinalities of tables affect the optimizer cost estimate Cardinality estimates are often poor quality Missing statistics Correlation between columns Complex operators Bad cardinality estimate results in poor plan

6. Problem – Query Variables select avg(e.salary), d.department_name from employees e, departments d where e.job_id = :job_v and e.employee_id = d.employee_id group by d.department_name Different bind values result in different intermediate cardinalities Optimal plan depends on bind value Plan chosen based on initial value then reused One plan not always appropriate for all bind values GB Y NLJ IX Optimal for “VP” GB Y HJ TS Optimal for “Sales Rep”

7. Solution

8. Solution Overview Two phase Solution 1.Collect information during execution 2.Use this information during subsequent optimization Cursor Cache SQL Execution Dictionary Parser Plan Generator Cost Estimator Query Transformation Code Generator SQL Compiler 1 2 User DB Application

9. Cursor Sharing Preliminaries A cursor stores the chosen plan in the cursor cache Caching avoids compilation overhead on every execution When a query is submitted “Soft parse” looks for matching cursor and uses it “Hard parse” no match found so query is compiled Prior to 11g Use bind variables – plan is shared Use literals – plan is not shared Optimizer peeks at bind value to estimate cardinalities

10. Cardinality Feedback

11. Addresses cardinality misestimates due to Missing statistics Inaccurate statistics Complex predicates Optimize the query as usual noting low quality estimates After execution compare actual cardinalities to estimates If estimates and actuals differ significantly Store actual values for later Force a re-optimization on next execution During optimization, use actuals instead of estimates

12. Example select e.employee_id, d.department_name from employees e, departments d where d.department_id = e.department_id and e.job_id = 'ST_CLERK' and e.salary < 3000 Difficult to estimate cardinality for equality AND inequality predicates First execution Estimate: 100 Actual: 1200 emp_job_ix dept_id_ix IX NL J Second execution Estimate: 1200 Actual: 1200 emp_job_ix departments TSTS IX HJ

13. Adaptive Cursor Sharing

14. Bind-Aware Cursor Sharing (BACS) Cursor sharing is aware of binds and selectivities Cursor stores a bind profile: Which binds’ values may affect plan Range of selectivities for each bind-containing predicate If a new bind value is outside of range, new cursor is built Cursors with same plan are merged Allowable selectivity range grows as cursors are merged 01 01 merge

15. Example select avg(e.salary), d.department_name from employees e, departments d where e.job_id = :job and e.department_id = d.department_id group by d.department_name; First execution: :job = VP sel(e.job_id = :job) = 2% GB Y NLJ dept_id_ix IX emp_job_ix IX 0 1 Second execution: :job = sales rep sel(e.job_id = :job) = 28% GB Y HJ employees TS departments TS 01

16. Example (continued) State of cursor cache after two executions: GB Y NLJ dept_id_ix IX emp_job_ix IX GB Y HJ emp TS dept TS 0 1 01 Third execution: job = clerk sel(e.job_id = :job) = 25% Use second cursor, selectivity is close enough

17. Adaptive Cursor Sharing Overhead of bind-aware cursor sharing Cursor matching is slower Additional re-optimizations consume CPU More cursors vying for space in cursor cache Use feedback to determine if BACS is warranted Adaptive cursor sharing (ACS) Monitor queries where BACS is applicable Identify queries with variable data volume Enable BACS for these queries

18. Example select avg(e.salary), d.department_name from employees e, departments d where e.job_id = :job and e.department_id = d.department_id group d.department_name; Big variation in # rows processed => use BACS First execution :job = VP 3 rows GB Y NLJ dept_id_ix IX emp_job_ix IX Second execution :job = sales rep 1000 rows GB Y NLJ dept_id_ix IX emp_job_ix IX

19. General Feedback Process no monitoring Feedback- enhanced Cursor Context Optimizer estimates Execution statistics … Cursor Context Optimizer estimates Execution statistics Enhanced compiler estimates… monitoring Cursor Context Optimizer estimates … SQL Compiler

20. Related Work Cardinality Feedback IBM Learning Optimizer (Stillger et al. VLDB 2001) Oracle Automatic SQL Tuning (Dageville et al. VLDB 2004) Adaptive Cursor Sharing SQL Server Auto Parameterization DB2 REOPT options

21. Experimental Results

22. Performance studies Oracle E-Business Suite (EBS) test database 60GB test system 14,000 tables, from tens to millions of rows EBS workload Over 300,000 queries Varying complexity, up to 140 tables per query Metrics Buffer gets Number of re-optimizations Number of distinct plans

23. Cardinality Feedback Execute 40,000 EBS queries twice each 10% of queries re-optimized, 4% get new plan

24. Performance Impact 924 queries improved by 68% 214 query degraded by 29% 58% overall improvement

25. Adaptive Cursor Sharing Approximately 400 queries average of 2.4 binds per query Queries known to generate multiple plans with different literals Machine-generated binds average of 55 bind sets per query 18% improvement with ACS 14% further improvement without cursor sharing

26. Results – Number of Plans Approximately 60 queries show multiple plans with ACS 20% fewer cursors with ACS vs. no sharing 45% more cursors with ACS vs. always sharing

27. Results – Impact of Feedback Compare ACS with and without feedback 5X fewer compilations with feedback 2X fewer cursors with feedback

28. Conclusions Feedback useful for two query optimization problems Intermediate cardinality estimation Optimization of queries with bind variables Solutions shipped in production release, Oracle 11g Performance improvement using a real SQL workload Future work Use framework for additional feedback-based optimization Persist feedback information

29. Questions & Answers

30. The preceding is intended to outline our general product direction. It is intended for information purposes only, and may not be incorporated into any contract. It is not a commitment to deliver any material, code, or functionality, and should not be relied upon in making purchasing decisions. The development, release, and timing of any features or functionality described for Oracle’s products remains at the sole discretion of Oracle.