Using adaptive cursor sharing (acs) to produce multiple optimal plans v2

Using Adaptive Cursor Sharing (ACS) to produce multiple Optimal Plans

Carlos Sierra

§ SQL Tuner handyman: developer, advisor, trainer, support §  IT: Enkitec (<0.2), Oracle(17), UNISYS(12), Ford(3), others(3) § Florida(17), Venezuela(3), Puerto Rico(6), Michigan(1), Mexico(x) § Tools: SQLTXPLAIN(SQLT), SQLHC, TRCANLZR(TRCA), others § Motto: Life’s Good!

Carlos Sierra Senior Technical Consultant at Enkitec

2

Adaptive Cursor Sharing (ACS)

§  Motivation

§  Mechanics

§  Test Case

§  Demo

§  Remarks

3

ACS Motivation

SQL Processing

§ Hard parse side effects –  CPU consumption –  Latch contention

§ Excessive hard parsing –  Affects concurrency –  Restricts scalability

§ Mitigating hard parsing –  Cursor sharing

Hard parsing is expensive!

5

Implementing Cursor Sharing Replacing literals with bind variables

6

Before Bind Peeking

§ Predicate –  WHERE channel_id = :b1

§ Unknowns –  Is :b1 between low and high values of channel_id? –  Is :b1 a popular value of channel_id? –  Are there any rows with value :b1 for channel_id?

§ Penalty –  Possible suboptimal plans

Before 9i CBO was blind to values passed

7

With Bind Peeking

§ Predicate –  WHERE channel_id = :b1

§ Plan is determined by peeked values –  EXEC :b1 := 9;

§ Optimal plan for 1st execution –  CBO can use low/high and histograms on channel_id

§ Penalty –  Possible suboptimal plans for subsequent executions on skewed data

9i offers a partial solution

8

Real-life Problem with Bind Peeking People Soft Payroll Application

9

§ WHERE employee BETWEEN :b1 AND :b2 § Payroll for one employee

–  :b1 = 123456 –  :b2 = 123456

§ Payroll for one company –  :b1 = 000001 –  :b2 = 999999

§ Doing payroll for an employee first then entire company

With Adaptive Cursor Sharing

§ Some queries are ACS candidates § Sophisticated non-persistent mechanism § Selectivity of predicates determine plan § Multiple optimal plans for a query!

–  If ACS is successfully applied § Penalty

–  Marginal increase in CPU and memory overhead

11g improves cursor sharing

10

ACS Mechanics

ACS high-level Overview High level overview

§  If SQL with binds meets some requirements –  Flag cursor as bind sensitive –  Start monitoring data volume manipulated by cursor

§  If bind sensitive and data volume manipulated by cursor varies significantly

–  Flag cursor as bind aware –  Start generating multiple optimal plans for this query on next hard parse

§  If bind aware then use selectivity of predicates to decide on plan

12

Bind Sensitive

§ SQL has explicit binds –  Or literals and cursor_sharing is “force”

§ Predicate: column + operand + bind_variable –  Equality operand “=“ and histogram on column

§  Ex: channel_id = :b1

–  Non-equality operand (range) regardless of histogram on column §  “>”, “>=“, “<“, ‘<=“, BETWEEN, LIKE

Minimum requirements

13

Bind Aware

§ Significant changes in data volume manipulated by cursor –  A few rows versus a few thousands of rows –  A few thousands of rows versus a few millions of rows

§ Specifying /*+ BIND_AWARE */ CBO Hint –  Bypasses the monitoring phase on data volume

How to become bind aware?

14

Plan Selection

§ Evaluate selectivity of predicates at soft parse § Compare to a non-persistent selectivity profile §  If within ranges of a known profile then select associated plan § Else hard parse

–  Compute and execute newly generated plan –  Create selectivity profile for new plan or update profile of existing plan

§  If ranges on selectivity profiles overlap then merge profiles

Based on selectivity profile of predicates

15

V$ dynamic views for ACS

§ V$SQL –  Shareable, bind sensitive and bind aware flags

§ V$SQL_CS_STATISTICS –  Data volume manipulated (rows processed)

§ V$SQL_CS_HISTOGRAM –  Record keeping of data volume per execution (small, medium, large)

§ V$SQL_CS_SELECTIVITY –  Predicates selectivity profiles

ACS non-persistent performance views

16

ACS Test Case

Our Query with Literals Guesstimate execution plan then verify it with demo 0

18

Possible Access Paths?

19

Optimal Execution Plan

20

Our Query with Bind Variables How many optimal execution plans can you foresee?

21

Multiple Optimal Plans for one Query Guesstimate optimal plan (access paths) for each query

Query :b1 :b2 AP1 AP2 q1 9 33 N1 N2 q2 5 32 q3 2 999 q4 9 999 q5 2 33

22

Multiple Optimal Plans for one Query Guesstimate optimal plan (access paths) for each query

Query :b1 :b2 AP1 AP2 q1 9 33 N1 N2 q2 5 32 N1 N2 q3 2 999 FTS FTS q4 9 999 N1 FTS q5 2 33 FTS N2

23

Multiple Optimal Plans for one Query Execute demos 1-5 and verify access paths

Query :b1 :b2 AP1 AP2 q1 9 33 N1 N2 q2 5 32 N1 N2 q3 2 999 FTS FTS q4 9 999 N1 FTS q5 2 33 FTS N2

24

Bind Sensitive: Rows Processed

§ Data volume manipulated –  Fuzzy representation –  S: few rows –  M: thousands or rows –  L: millions of rows

§  v$sql_cs_histogram –  Bucket(0): S –  Bucket(1): M –  Bucket(2): L

Monitor v$sql_cs_statistics.rows_processed

Query :b1 :b2 Optimal Rows

Processed q1 9 33 N1/N2 37,382 q2 5 32 N1/N2 2 q3 2 999 FTS/FTS 8,021,324 q4 9 999 N1/FTS 6,233,815 q5 2 33 FTS/N2 1,825,131

25

ACS Demo

Demo 6: When Cursor becomes Bind Aware?


Processed Bucket Aware Child Actual q1 9 33 N1/N2 37,382 q2 5 32 N1/N2 2 q3 2 999 FTS/FTS 8,021,324 q4 9 999 N1/FTS 6,233,815 q5 2 33 FTS/N2 1,825,131

Obtain rows processed from demo 1-5 then guesstimate aware flag

27



Processed Bucket Aware Child Actual q1 9 33 N1/N2 37,382 1 q2 5 32 N1/N2 2 0 q3 2 999 FTS/FTS 8,021,324 2 q4 9 999 N1/FTS 6,233,815 2 q5 2 33 FTS/N2 1,825,131 2


28



Processed Bucket Aware Child Actual q1 9 33 N1/N2 37,382 1 N q2 5 32 N1/N2 2 0 N q3 2 999 FTS/FTS 8,021,324 2 Y q4 9 999 N1/FTS 6,233,815 2 Y q5 2 33 FTS/N2 1,825,131 2 Y


29



Processed Bucket Aware Child Actual q1 9 33 N1/N2 37,382 1 N 0 N1/N2 q2 5 32 N1/N2 2 0 N 0 N1/N2 q3 2 999 FTS/FTS 8,021,324 2 Y 1 FTS/FTS q4 9 999 N1/FTS 6,233,815 2 Y 2 N1/FTS q5 2 33 FTS/N2 1,825,131 2 Y 3 FTS/N2


30



Processed Bucket Aware Child Actual q1 9 33 N1/N2 37,382 1 N 0 N1/N2 q2 5 32 N1/N2 2 0 N 0 N1/N2 q3 2 999 FTS/FTS 8,021,324 2 Y 1 FTS/FTS q4 9 999 N1/FTS 6,233,815 2 Y 2 N1/FTS q5 2 33 FTS/N2 1,825,131 2 Y 3 FTS/N2


31



Processed Bucket Aware Child Actual q5 2 33 FTS/N2 1,825,131 q4 9 999 N1/FTS 6,233,815 q3 2 999 FTS/FTS 8,021,324 q2 5 32 N1/N2 2 q1 9 33 N1/N2 37,382

Compute bucket and guesstimate aware flag and actual plan

32



Processed Bucket Aware Child Actual q5 2 33 FTS/N2 1,825,131 2 q4 9 999 N1/FTS 6,233,815 2 q3 2 999 FTS/FTS 8,021,324 2 q2 5 32 N1/N2 2 0 q1 9 33 N1/N2 37,382 1


33



Processed Bucket Aware Child Actual q5 2 33 FTS/N2 1,825,131 2 N q4 9 999 N1/FTS 6,233,815 2 N q3 2 999 FTS/FTS 8,021,324 2 N q2 5 32 N1/N2 2 0 N q1 9 33 N1/N2 37,382 1 Y


34



Processed Bucket Aware Child Actual q5 2 33 FTS/N2 1,825,131 2 N 0 FTS/N2 q4 9 999 N1/FTS 6,233,815 2 N 0 FTS/N2 q3 2 999 FTS/FTS 8,021,324 2 N 0 FTS/N2 q2 5 32 N1/N2 2 0 N 0 FTS/N2 q1 9 33 N1/N2 37,382 1 Y 1 N1/N2


35



Processed Bucket Aware Child Actual q5 2 33 FTS/N2 1,825,131 2 N 0 FTS/N2 q4 9 999 N1/FTS 6,233,815 2 N 0 FTS/N2 q3 2 999 FTS/FTS 8,021,324 2 N 0 FTS/N2 q2 5 32 N1/N2 2 0 N 0 FTS/N2 q1 9 33 N1/N2 37,382 1 Y 1 N1/N2


36



Processed Bucket Aware Child Actual q5 2 33 FTS/N2 1,825,131 2 q4 9 999 N1/FTS 6,233,815 2 q3 2 999 FTS/FTS 8,021,324 2 q1 9 33 N1/N2 37,382 1 q2 5 32 N1/N2 2 0

Guesstimate aware flag and actual plan

37



Processed Bucket Aware Child Actual q5 2 33 FTS/N2 1,825,131 2 N q4 9 999 N1/FTS 6,233,815 2 N q3 2 999 FTS/FTS 8,021,324 2 N q1 9 33 N1/N2 37,382 1 N q2 5 32 N1/N2 2 0 N


38



Processed Bucket Aware Child Actual q5 2 33 FTS/N2 1,825,131 2 N 0 FTS/N2 q4 9 999 N1/FTS 6,233,815 2 N 0 FTS/N2 q3 2 999 FTS/FTS 8,021,324 2 N 0 FTS/N2 q1 9 33 N1/N2 37,382 1 N 0 FTS/N2 q2 5 32 N1/N2 2 0 N 0 FTS/N2


39



Processed Bucket Aware Child Actual q5 2 33 FTS/N2 1,825,131 2 N 0 FTS/N2 q4 9 999 N1/FTS 6,233,815 2 N 0 FTS/N2 q3 2 999 FTS/FTS 8,021,324 2 N 0 FTS/N2 q1 9 33 N1/N2 37,382 1 N 0 FTS/N2 q2 5 32 N1/N2 2 0 N 0 FTS/N2


40

Real-life Problem with ACS People Soft Payroll Application

41

§ WHERE employee BETWEEN :b1 AND :b2 § Payroll for one employee

–  :b1 = 123456 –  :b2 = 123456

§ Payroll for one company –  :b1 = 000001 –  :b2 = 999999

§ Doing payroll for a few employees first then entire company

Closing Remarks

Understanding Selectivity Profile From demo 6

Query :b1 :b2 Child q3 2 999 1 q4 9 999 2 q5 2 33 3

43



44



45

Remarks on Bind Sensitivity

§ Monitor V$SQL_CS_STATISTICS.rows_processed –  If small number of rows then

§  V$SQL_CS_HISTOGRAM.bucket_id(0)++

–  If medium number of rows then §  V$SQL_CS_HISTOGRAM.bucket_id(1)++

–  If large number of rows then §  V$SQL_CS_HISTOGRAM.bucket_id(2)++

Based on experimental observation

46

Remarks on Bind Aware

§ Some cases where cursor may become bind aware –  bucket_id(0) = bucket_id(1) > 0 –  bucket_id(1) = bucket_id(2) > 0 –  bucket_id(0) > 0 and bucket_id(2) > 0

§ Or use /*+ BIND_AWARE */ CBO Hint –  What if we cannot modify code?

Based on experimental observation

47

Conclusions

§ ACS only applies to a subset of queries with binds § ACS requires a ramp-up process (few executions) §  In some cases cursor may fail to become bind aware § To force a cursor become bind aware use CBO Hint § ACS is not persistent § ACS works well with SQL Plan Management

ACS can produce multiple optimal plans for one query

48

Give Away

§ Creates a SQL Patch for one SQL_ID § Turns “on” EVENT 10053 for SQL_ID § Hints on SQL Patch

–  GATHER_PLAN_STATISTICS –  MONITOR –  BIND_AWARE

§ Consider using and customizing this free script

Script sqlt/utl/coe_gen_sql_patch.sql (MOS 215187.1)

49

References and Contact Info

§ https://blogs.oracle.com/optimizer/ –  Insight into the workings of the Optimizer

Oracle Optimizer Blog

§  [email protected] § http://carlos-sierra.net § @csierra_usa

50

Using adaptive cursor sharing (acs) to produce multiple optimal plans v2

Technology