Oracle中為了提高sql的執行效率,需要減少硬解析,實作shared cursor共享,最常見的方法是使用綁定變量,但很多時候由于各種原因未能在開發初期使用綁定變量,對于減少硬解析的目的,退而求其次地方法是設定cursor_sharing.
1.準備環境
實驗環境
[email protected]> select * from v$version;
BANNER
----------------------------------------------------------------
Oracle Database 10g Enterprise Edition Release 10.2.0.1.0 - Prod
PL/SQL Release 10.2.0.1.0 - Production
CORE 10.2.0.1.0 Production
TNS for Linux: Version 10.2.0.1.0 - Production
NLSRTL Version 10.2.0.1.0 – Production
使用腳本插入資料後:
[email protected]> desc t;
Name Null? Type
----------------------- -------- ----------------
ID VARCHAR2(5)
NAME NUMBER(38)
BALLO[email protected]> select id,count(*) from t group by id;
ID COUNT(*)
----- ----------
d 6
a 10000
b 20000
c 20
[email protected]> create index ind_id on t(id);
Index created.
2.取值為exact時(預設):
[email protected]> show parameter cursor_sharing;
NAME TYPE VALUE
------------------------------------ ----------- -----------------
cursor_sharing string EXACT
[email protected]> select count(*) from t where id='b';
COUNT(*)
----------
20000
[email protected]> select count(*) from t where id='d';
COUNT(*)
----------
6
[email protected]> select sql_text,sql_id,version_count,executions from v$sqlarea
where sql_text like 'select count(*) from t where id=%';
SQL_TEXT SQL_ID VERSION_COUNT EXECUTIONS
---------------------------------------- ------------- ------------- ----------
select count(*) from t where id='d' b0gfs7u9r55rv 1 1
select count(*) from t where id='b' fqurbumy7bsg6 1 1
可以看到兩條查詢語句沒有使用綁定變量,有各自對應的sql_id,子遊标數均為1個。兩個sql查詢沒有任何關系。
檢視兩次sql的執行計劃:
[email protected]>select * from table(dbms_xplan.
display_cursor('b0gfs7u9r55rv',0,'advanced'));
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------
SQL_ID b0gfs7u9r55rv, child number 0
-------------------------------------
select count(*) from t where id='d'
Plan hash value: 3666266488
---------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
----------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | | | 1 (100)| |
| 1 | SORT AGGREGATE | | 1 | 2 | | |
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------
|* 2 | INDEX RANGE SCAN| IND_ID | 14 | 28 | 1 (0)| 00:00:01 |
----------------------------------------------------------------------------
第一次查詢利用了索引。
[email protected]> select * from table(dbms_xplan.display_cursor('fqurbumy7bsg6',0,'advanced'));
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------
SQL_ID fqurbumy7bsg6, child number 0
-------------------------------------
select count(*) from t where id='b'
Plan hash value: 2966233522
---------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
---------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | | | 15 (100)| |
| 1 | SORT AGGREGATE | | 1 | 2 | | |
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------
|* 2 | TABLE ACCESS FULL| T | 19783 | 39566 | 15 (0)| 00:00:01 |
---------------------------------------------------------------------------
第二次查詢利用了全表掃描
我們在更直覺地來看一下兩次sql查詢後的硬解析統計情況:
[email protected]> select count(*) from t where id='a';
COUNT(*)
----------
10000
[email protected]> select name,value from v$sysstat where name like '%parse%';
NAME VALUE
---------------------------------------------------------------- ----------
parse time cpu 2133
parse time elapsed 4463
parse count (total) 54889
parse count (hard) 6579(硬解析數目)
parse count (failures) 52
[email protected]> select count(*) from t where id='c';
COUNT(*)
----------
20
[email protected]> select name,value from v$sysstat where name like '%parse%';
NAME VALUE
---------------------------------------------------------------- ----------
parse time cpu 2134
parse time elapsed 4464
parse count (total) 54895
parse count (hard) 6580(硬解析數目)
parse count (failures) 52
硬解析數目再次加1
因為資料的巨大差異性,導緻了對兩次查詢有不同的執行計劃,這也說明在cursor設定為exact時,兩條sql語句如果存在一點不同,就不會共享cursor,而進行兩次硬解析。
3.設定為force時
Oracle對輸入的SQL值,會将where條件取值自動替換為綁定變量。以後在輸入相同的結構SQL語句時,會進行cursor sharing共享遊标
[email protected]> alter system set cursor_sharing=force;
System altered.
[email protected]> show parameter cursor_sharing;
NAME TYPE VALUE
------------------------- ------------ -------
cursor_sharing string FORCE
清除一下share pool中已緩存的cursor
[email protected]> alter system flush shared_pool;
System altered.
檢視硬解析情況:
[email protected]> select count(*) from t where id='b';
COUNT(*)
----------
20000
[email protected]> select name,value from v$sysstat where name like '%parse%';
NAME VALUE
------------------------------------- ----------
parse time cpu 2163
parse time elapsed 4506
parse count (total) 55097
parse count (hard) 6668
parse count (failures) 52
[email protected]> select count(*) from t where id='d';
COUNT(*)
----------
6
[email protected]> select name,value from v$sysstat where name like '%parse%';
NAME VALUE
----------------------------------------- ----------
parse time cpu 2164
parse time elapsed 4507
parse count (total) 55101
parse count (hard) 6669
parse count (failures) 52
硬解析加1了,這不應該呀!!
[email protected]> select sql_text,child_number from v$sql where sql_text like 'select count(*) from t where id%';
SQL_TEXT CHILD_NUMBER
---------------------------------------- ------------
select count(*) from t where id='d' 0
select count(*) from t where id='b' 0
可以看到并沒有使用綁定變量,force的設定沒有生效。
解決辦法:應在設定cursor_sharing前,執行兩次下面語句:
alter system flush shared_pool;
[email protected]> alter system flush shared_pool;
System altered.
[email protected]> alter system flush shared_pool;
System altered.
[email protected]> alter system set cursor_sharing=force;
System altered.
設定好了,接着進行sql測試
[email protected]> select count(*) from t where id='d';
COUNT(*)
----------
6
[email protected]> select name,value from v$sysstat where name like '%parse%';
NAME VALUE
-------------------------------------- ----------
parse time cpu 2216
parse time elapsed 4572
parse count (total) 55867
parse count (hard) 6910
parse count (failures) 55
[email protected]> select count(*) from t where id='b';
COUNT(*)
----------
20000
[email protected]> select name,value from v$sysstat where name like '%parse%';
NAME VALUE
--------------------------------------- ----------
parse time cpu 2216
parse time elapsed 4572
parse count (total) 55869
parse count (hard) 6910
parse count (failures) 55
硬解析的次數沒有發生變化
[email protected]> select sql_text,sql_id,version_count,executions from v$sqlarea
2 where sql_text like 'select count(*) from t where%';
SQL_TEXT SQL_ID VERSION_COUN EXECUTIONS
---------------------------------------------- ------------- ------------- -------------
select count(*) from t where id=:"SYS_B_0" g82ztj8p3q174 1 2
可以看到兩次查詢使用了綁定變量,将謂詞的值用sys_B_0代替。該語句執行了兩次,有一個child cursor(子遊标)。
在來看一下兩次查詢語句的執行計劃:
[email protected]> select * from table(dbms_xplan.display_cursor('g82ztj8p3q174',0,'advanced'));
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------
SQL_ID g82ztj8p3q174, child number 0
-------------------------------------
select count(*) from t where id=:"SYS_B_0"
Plan hash value: 3666266488
----------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
----------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | | | 1 (100)| |
| 1 | SORT AGGREGATE | | 1 | 4 | | |
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------
|* 2 | INDEX RANGE SCAN| IND_ID | 6 | 24 | 1 (0)| 00:00:01 |
----------------------------------------------------------------------------
兩次的查詢使用了同一個執行計劃:索引掃描。這就是force的設定。
對與參數cusor_sharing設定為force時,根據實驗,我們可以得出下列結論:
- Oracle對輸入的SQL值,會将where條件取值自動替換為綁定變量。以後在輸入相同的結構SQL語句時,會進行cursor sharing共享遊标;
- 在第一次進行自動替換綁定變量的時候,Oracle會依據bind peeking取值,擷取到一個執行計劃,對應成子遊标;
- 在以後的SQL語句中,隻要出現父遊标可共享的情況,Oracle都會強制使用生成的唯一子遊标進行sharing。不去在乎是不是對SQL是最優的執行計劃;
- FORCE取值的規則思想很簡單,對SQL語句進行強制的綁定變量替換。使用第一次的bind peeking值生成執行計劃,之後全部使用這個執行計劃。這種方式實作了遊标共享,避免出現大量的library cache硬解析,限制一個父遊标的version_count數量。
- 如果這種SQL語句本身是“Good SQL”,也就是條件列分布比較平均,沒有出現過大的偏移分布。我們認為這種FORCE是很有益的。但是如果資料列分布不平均,這樣借用第一次輸入的bind peeking生成并且共享執行計劃就很成問題。我們說,在cursor_sharing取定FORCE遇到的潛在問題,和我們使用綁定變量時候使用的bind peeking值問題是相同的。
4.設定為similar時
當cursor_sharing設定為SIMILAR的時候,Oracle對沒有使用綁定變量的SQL字面語句都會進行處理,将where條件後自動替換為綁定變量;并在執行語句時,對每一個條件設定值,都會生成一個新的child cursor子遊标,比對一個新的執行計劃。
同上面的force設定:
[email protected]> alter system flush shared_pool;
System altered.
[email protected]> alter system flush shared_pool;
System altered.
[email protected]> alter system set cursor_sharing=similar;
System altered.
[email protected]> show parameter cursor_sharing;
NAME TYPE VALUE
------------------------------------ ------------------------------- -----------------------
cursor_sharing string SIMILAR
收集一下統計資訊和直方圖:
[email protected]> exec dbms_stats.gather_table_stats('BALLONTT','T', cascade => true,METHOD_OPT =>'FOR COLUMNS SIZE AUTO');
PL/SQL procedure successfully completed.
[email protected]> select column_name, num_distinct, NUM_BUCKETS, HISTOGRAM from dba_tab_col_statistics
2 where wner='BALLONTT' and table_name='T';
COLUMN_NAME NUM_DISTINCT NUM_BUCKETS HISTOGRAM
------------------------------ ------------ ----------- ---------------
ID 4 4 FREQUENCY
NAME 10000 1 NONE
進行測試:
[email protected]> select count(*) from t where id='d';
COUNT(*)
----------
6
[email protected]> select name,value from v$sysstat where name like '%parse%';
NAME VALUE
------------------------------------------- ----------
parse time cpu 2412
parse time elapsed 4792
parse count (total) 62015
parse count (hard) 8059
parse count (failures) 64
[email protected]> select count(*) from t where id='b';
COUNT(*)
----------
20000
[email protected]> select name,value from v$sysstat where name like '%parse%';
NAME VALUE
----------------------------------------- ----------
parse time cpu 2412
parse time elapsed 4792
parse count (total) 62017
parse count (hard) 8060
parse count (failures) 64
硬解析次數加1
[email protected]> select sql_text,sql_id,child_number,executions from v$sql where sql_text like 'select count(*) from t where id%';
SQL_TEXT SQL_ID CHILD_NUMBER EXECUTIONS
----------------------------------------- ---------------- ------------- ------------
select count(*) from t where id=:"SYS_B_0" g82ztj8p3q174 0 1
select count(*) from t where id=:"SYS_B_0" g82ztj8p3q174 1 1
[email protected]> select sql_text,sql_id,version_count,executions from v$sqlarea where sql_Text like 'select count(*) from t where id%';
SQL_TEXT SQL_ID VERSION_COUNT EXECUTIONS
----------------------------------------------- ------------- ------------- ---------
select count(*) from t where id=:"SYS_B_0" g82ztj8p3q174 2 2
使用了綁定變量,但是兩次查詢生成了兩個子遊标(child cursor),即對應兩個執行計劃。
為了更清楚地看到兩次sql使用了兩個不同的子遊标,即對應着使用了兩個執行計劃,我們再來分步查詢一次:
[email protected]> select count(*) from t where id='d';
COUNT(*)
----------
6
[email protected]> select sql_text,sql_id,child_number from v$sql where sql_text like '%select count(*) from t where id%';
SQL_TEXT SQL_ID CHILD_NUMBER
----------------------------------------------- ------------- -------------
select count(*) from t where id=:"SYS_B_0" g82ztj8p3q174 0
檢視執行計劃:
[email protected]> select * from table(dbms_xplan.display_cursor('g82ztj8p3q174',0,'advanced'));
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------
SQL_ID g82ztj8p3q174, child number 0
-------------------------------------
select count(*) from t where id=:"SYS_B_0"
Plan hash value: 3666266488
----------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
----------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | | | 1 (100)| |
| 1 | SORT AGGREGATE | | 1 | 2 | | |
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------
|* 2 | INDEX RANGE SCAN| IND_ID | 1 | 2 | 1 (0)| 00:00:01 |
----------------------------------------------------------------------------
執行計劃為索引路徑
再次查詢:
[email protected]> select count(*) from t where id='b';
COUNT(*)
----------
20000
[email protected]> select sql_text,sql_id,child_number from v$sql where sql_text like 'select count(*) from t where id%'
2 ;
SQL_TEXT SQL_ID CHILD_NUMBER
----------------------------------------------- ------------- -------------
select count(*) from t where id=:"SYS_B_0" g82ztj8p3q174 0
select count(*) from t where id=:"SYS_B_0" g82ztj8p3q174 1
檢視子遊标(chilid_number=1)的執行計劃:
[email protected]> select * from table(dbms_xplan.display_cursor('g82ztj8p3q174',1,'advanced'));
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------
SQL_ID g82ztj8p3q174, child number 1
-------------------------------------
select count(*) from t where id=:"SYS_B_0"
Plan hash value: 2966233522
---------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
---------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | | | 15 (100)| |
| 1 | SORT AGGREGATE | | 1 | 2 | | |
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------
|* 2 | TABLE ACCESS FULL| T | 20081 | 40162 | 15 (0)| 00:00:01 |
---------------------------------------------------------------------------
執行計劃為全表掃描,生成了一個新的執行計劃。
由此我們可以得到結論:
- 當cursor_sharing設定為SIMILAR的時候,Oracle對沒有使用綁定變量的SQL字面語句都會進行處理,将where條件後自動替換為綁定變量;
- 在執行語句是,對每一個條件設定值,都會生成一個新的child cursor子遊标,與父遊标相對應。當一個語句輸入的時候,如果之前存在過相同條件值的SQL子遊标,就共享該子遊标。否則生成一個新的child cursor,生成一個比對的執行計劃;
- SIMILAR是FORCE的進化版。在SIMILAR模式下,Oracle對遊标共享的條件變得比較敏感。如果綁定變量值發生變化,就意味着執行計劃可能存在不比對的情況。是以索性Oracle對每一個新的值都bind peeking一下,生成執行計劃。而執行計劃遊标的共享隻在相同綁定變量的時候才發生。
- 這個與FORCE相比,cursor_sharing=SIMILAR的确緩解了由于bind peeking單次帶來的執行計劃不比對問題。但是會引入兩個新問題:
- 如果對應條件列的取值相對較少,這樣生成執行計劃的個數起碼是可以控制的。如 果是一種連續取值情況或者對應取值很多,必然引起parent cursor對應的child cursor數目增多,每次從child cursor列中周遊的時間增加,latch和pin發生的時間增多。這也是similar取值是一個常見的問題;
- 生成child cursor的标準不是是否執行計劃相同,而是綁定變量值相同。這樣如果資料分布較為平均,所有值對應的執行計劃都是相同的。那麼生成很多的子遊标執行計劃必然是相同的。這樣又會帶來性能和其他一些問題。
- 在謂詞條件的列後沒有直方圖時,其作用和force類似,會共享子遊标(可以删除where條件後列上的直方圖重複最後一步實驗驗證)
Cursor_sharing設定為similar,存在一些BUG,即使是force也同樣有BUG存在。例如導緻EXPDP的導出時間變成,物化視圖的更新變慢等。Oracle 并不建議你設定此參數。在12C中ORACLE将廢除similar的設定。
ballontt
2013/12/25
---The End---
微網誌:weibo.com/ballontt 如需轉載,請标明出處和連結,謝謝!