一次关于SQL优化的分析全过程 [转帖]_MySQL, Oracle及数据库讨论区_Weblogic技术|Tuxedo技术|中间件技术|Oracle论坛|JAVA论坛|Linux/Unix技术|hadoop论坛

总帖数

每页帖数

1/1页

返回列表

发起投票

查看: 3642 | 回复: 0

主题： 一次关于SQL优化的分析全过程 [转帖]

Gavin

注册用户

等级：少校
经验：878
发帖：130
精华：1
注册：2011-7-21
状态：离线
发送短消息息给Gavin

加好友发送短消息息给Gavin

发消息

发表于：

2011-8-2 9:11:23 | [全部帖] [楼主帖]

楼主

转自：http://www.itpub.net/viewthread.php?tid=91642&extra=page%3D12%26amp%3Bfilter%3Ddigest

一次分析的全过程,和大家交流!

我们的程序员抱怨一段执行很慢的程序，我把代码执行分析执行计划，

后果如下：

这是最初的执行效果及执行计划

SQL> SELECT "SP_TRANS"."TRANS_NO",
2 "SP_TRANS_SUB"."ITEM_CODE",
3 "SP_ITEM"."ITEM_NAME",
4 "SP_ITEM"."CHART_ID",
5 "SP_ITEM"."SPECIFICATION",
6 "SP_TRANS_SUB"."COUNTRY",
7 "SP_TRANS_SUB"."QTY",
8 "SP_TRANS_SUB"."PRICE",
9 "SP_TRANS"."VENDOR_CODE",
10 "SP_TRANS"."PAY_MODE",
11 NVL("SP_TRANS_SUB"."PAY_QTY",0),
12 0 as PAY_THIS
13 FROM "SP_ITEM",
14 "SP_TRANS_SUB",
15 "SP_TRANS"
16 WHERE ( "SP_TRANS_SUB"."TRANS_NO" = "SP_TRANS"."TRANS_NO" ) and
17 ( "SP_ITEM"."ITEM_CODE" = "SP_TRANS_SUB"."ITEM_CODE" ) and
18 ( ( "SP_TRANS"."VENDOR_CODE" = '20011021023') )
19 /
8 rows selected.
Elapsed: 00: 00: 00.51
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE
1 0 NESTED LOOPS
2 1 NESTED LOOPS
3 2 TABLE ACCESS (FULL) OF 'SP_TRANS'
4 2 TABLE ACCESS (BY INDEX ROWID) OF 'SP_TRANS_SUB'
5 4 INDEX (RANGE SCAN) OF 'PK_SP_TRANS_SUB' (UNIQUE)
6 1 TABLE ACCESS (BY INDEX ROWID) OF 'SP_ITEM'
7 6 INDEX (UNIQUE SCAN) OF 'PK_SP_ITEM' (UNIQUE)
Statistics
----------------------------------------------------------
0 recursive calls
4 db block gets
323 consistent gets
0 physical reads
0 redo size
1809 bytes sent via SQL*Net to client
425 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
8 rows processed

此前这几个表都没有分析过。

然后我analyze相关表

SQL> analyze table sp_trans_sub compute statistics;
Table analyzed.
Elapsed: 00: 00: 30.64
SQL> SELECT "SP_TRANS"."TRANS_NO",
2 "SP_TRANS_SUB"."ITEM_CODE",
3 "SP_ITEM"."ITEM_NAME",
4 "SP_ITEM"."CHART_ID",
5 "SP_ITEM"."SPECIFICATION",
6 "SP_TRANS_SUB"."COUNTRY",
7 "SP_TRANS_SUB"."QTY",
8 "SP_TRANS_SUB"."PRICE",
9 "SP_TRANS"."VENDOR_CODE",
10 "SP_TRANS"."PAY_MODE",
11 NVL("SP_TRANS_SUB"."PAY_QTY",0),
12 0 as PAY_THIS
13 FROM "SP_ITEM",
14 "SP_TRANS_SUB",
15 "SP_TRANS"
16 WHERE ( "SP_TRANS_SUB"."TRANS_NO" = "SP_TRANS"."TRANS_NO" ) and
17 ( "SP_ITEM"."ITEM_CODE" = "SP_TRANS_SUB"."ITEM_CODE" ) and
18 ( ( "SP_TRANS"."VENDOR_CODE" = '20011021023') )
19 /
8 rows selected.
Elapsed: 00: 00: 06.49
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=18577 Card=126726520
Bytes=30034185240)
1 0 MERGE JOIN (Cost=18577 Card=126726520 Bytes=30034185240)
2 1 SORT (JOIN) (Cost=14722 Card=310300 Bytes=20790100)
3 2 HASH JOIN (Cost=358 Card=310300 Bytes=20790100)
4 3 TABLE ACCESS (FULL) OF 'SP_TRANS' (Cost=43 Card=229
Bytes=8473)
5 3 TABLE ACCESS (FULL) OF 'SP_TRANS_SUB' (Cost=158 Card
=135502 Bytes=4065060)
6 1 SORT (JOIN) (Cost=3855 Card=40840 Bytes=6942800)
7 6 TABLE ACCESS (FULL) OF 'SP_ITEM' (Cost=77 Card=40840 B
ytes=6942800)
Statistics
----------------------------------------------------------
150 recursive calls
89 db block gets
1837 consistent gets
755 physical reads
60 redo size
1732 bytes sent via SQL*Net to client
425 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
4 sorts (memory)
1 sorts (disk)
8 rows processed
SQL>
SQL> analyze table sp_trans compute statistics;
Table analyzed.
Elapsed: 00: 00: 13.00
SQL>
SQL> SELECT "SP_TRANS"."TRANS_NO",
2 "SP_TRANS_SUB"."ITEM_CODE",
3 "SP_ITEM"."ITEM_NAME",
4 "SP_ITEM"."CHART_ID",
5 "SP_ITEM"."SPECIFICATION",
6 "SP_TRANS_SUB"."COUNTRY",
7 "SP_TRANS_SUB"."QTY",
8 "SP_TRANS_SUB"."PRICE",
9 "SP_TRANS"."VENDOR_CODE",
10 "SP_TRANS"."PAY_MODE",
11 NVL("SP_TRANS_SUB"."PAY_QTY",0),
12 0 as PAY_THIS
13 FROM "SP_ITEM",
14 "SP_TRANS_SUB",
15 "SP_TRANS"
16 WHERE ( "SP_TRANS_SUB"."TRANS_NO" = "SP_TRANS"."TRANS_NO" ) and
17 ( "SP_ITEM"."ITEM_CODE" = "SP_TRANS_SUB"."ITEM_CODE" ) and
18 ( ( "SP_TRANS"."VENDOR_CODE" = '20011021023') )
19 /
8 rows selected.
Elapsed: 00: 00: 01.62
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=1453 Card=447198 Byt
es=101066748)
1 0 NESTED LOOPS (Cost=1453 Card=447198 Bytes=101066748)
2 1 HASH JOIN (Cost=358 Card=1095 Bytes=61320)
3 2 TABLE ACCESS (FULL) OF 'SP_TRANS' (Cost=43 Card=273 By
tes=7098)
4 2 TABLE ACCESS (FULL) OF 'SP_TRANS_SUB' (Cost=158 Card=1
35502 Bytes=4065060)
5 1 TABLE ACCESS (BY INDEX ROWID) OF 'SP_ITEM' (Cost=1 Card=
40840 Bytes=6942800)
6 5 INDEX (UNIQUE SCAN) OF 'PK_SP_ITEM' (UNIQUE)
Statistics
----------------------------------------------------------
0 recursive calls
8 db block gets
1344 consistent gets
0 physical reads
0 redo size
1824 bytes sent via SQL*Net to client
425 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
3 sorts (memory)
0 sorts (disk)
8 rows processed
SQL>
SQL> analyze table sp_item compute statistics
2 /
Table analyzed.
Elapsed: 00: 00: 11.67
SQL> SELECT "SP_TRANS"."TRANS_NO",
2 "SP_TRANS_SUB"."ITEM_CODE",
3 "SP_ITEM"."ITEM_NAME",
4 "SP_ITEM"."CHART_ID",
5 "SP_ITEM"."SPECIFICATION",
6 "SP_TRANS_SUB"."COUNTRY",
7 "SP_TRANS_SUB"."QTY",
8 "SP_TRANS_SUB"."PRICE",
9 "SP_TRANS"."VENDOR_CODE",
10 "SP_TRANS"."PAY_MODE",
11 NVL("SP_TRANS_SUB"."PAY_QTY",0),
12 0 as PAY_THIS
13 FROM "SP_ITEM",
14 "SP_TRANS_SUB",
15 "SP_TRANS"
16 WHERE ( "SP_TRANS_SUB"."TRANS_NO" = "SP_TRANS"."TRANS_NO" ) and
17 ( "SP_ITEM"."ITEM_CODE" = "SP_TRANS_SUB"."ITEM_CODE" ) and
18 ( ( "SP_TRANS"."VENDOR_CODE" = '20011021023') )
19 /
8 rows selected.
Elapsed: 00: 00: 01.43
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=687 Card=1362 Bytes=
128028)
1 0 HASH JOIN (Cost=687 Card=1362 Bytes=128028)
2 1 HASH JOIN (Cost=358 Card=1362 Bytes=76272)
3 2 TABLE ACCESS (FULL) OF 'SP_TRANS' (Cost=43 Card=273 By
tes=7098)
4 2 TABLE ACCESS (FULL) OF 'SP_TRANS_SUB' (Cost=158 Card=1
35502 Bytes=4065060)
5 1 TABLE ACCESS (FULL) OF 'SP_ITEM' (Cost=77 Card=29547 Byt
es=1122786)
Statistics
----------------------------------------------------------
0 recursive calls
12 db block gets
1820 consistent gets
0 physical reads
0 redo size
1732 bytes sent via SQL*Net to client
425 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
3 sorts (memory)
0 sorts (disk)
8 rows processed

大家可以看出执行计划是怎样一步一步的改变的。

最后Oracle给我选择了三个全表扫描。而且，从db block gets 、consistent gets 全都“提高”了。

然后我analyze索引

SQL> analyze table sp_trans compute statistics for all indexes;

表已分析。

已用时间: 00: 00: 01.12

已选择8行。

已用时间: 00: 00: 01.82

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=687 Card=1362 Bytes=
128028)
1 0 HASH JOIN (Cost=687 Card=1362 Bytes=128028)
2 1 HASH JOIN (Cost=358 Card=1362 Bytes=76272)
3 2 TABLE ACCESS (FULL) OF 'SP_TRANS' (Cost=43 Card=273 By
tes=7098)
4 2 TABLE ACCESS (FULL) OF 'SP_TRANS_SUB' (Cost=158 Card=1
35502 Bytes=4065060)
5 1 TABLE ACCESS (FULL) OF 'SP_ITEM' (Cost=77 Card=29547 Byt
es=1122786)
Statistics
----------------------------------------------------------
72 recursive calls
12 db block gets
1851 consistent gets
0 physical reads
0 redo size
1732 bytes sent via SQL*Net to client
425 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
3 sorts (memory)
0 sorts (disk)
8 rows processed
SQL> analyze table sp_trans_sub compute statistics for all indexes;

表已分析。

已用时间: 00: 00: 05.67

已选择8行。

已用时间: 00: 00: 01.82

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=687 Card=1362 Bytes=
128028)
1 0 HASH JOIN (Cost=687 Card=1362 Bytes=128028)
2 1 HASH JOIN (Cost=358 Card=1362 Bytes=76272)
3 2 TABLE ACCESS (FULL) OF 'SP_TRANS' (Cost=43 Card=273 By
tes=7098)
4 2 TABLE ACCESS (FULL) OF 'SP_TRANS_SUB' (Cost=158 Card=1
35502 Bytes=4065060)
5 1 TABLE ACCESS (FULL) OF 'SP_ITEM' (Cost=77 Card=29547 Byt
es=1122786)
Statistics
----------------------------------------------------------
0 recursive calls
12 db block gets
1820 consistent gets
0 physical reads
0 redo size
1732 bytes sent via SQL*Net to client
425 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
3 sorts (memory)
0 sorts (disk)
8 rows processed
SQL> analyze table sp_item compute statistics for all indexes
2 /

表已分析。

已用时间: 00: 00: 01.12

已选择8行。

已用时间: 00: 00: 01.82

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=687 Card=1362 Bytes=
128028)
1 0 HASH JOIN (Cost=687 Card=1362 Bytes=128028)
2 1 HASH JOIN (Cost=358 Card=1362 Bytes=76272)
3 2 TABLE ACCESS (FULL) OF 'SP_TRANS' (Cost=43 Card=273 By
tes=7098)
4 2 TABLE ACCESS (FULL) OF 'SP_TRANS_SUB' (Cost=158 Card=1
35502 Bytes=4065060)
5 1 TABLE ACCESS (FULL) OF 'SP_ITEM' (Cost=77 Card=29547 Byt
es=1122786)
Statistics
----------------------------------------------------------
0 recursive calls
12 db block gets
1820 consistent gets
0 physical reads
0 redo size
1732 bytes sent via SQL*Net to client
425 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
3 sorts (memory)
0 sorts (disk)
8 rows processed
SQL>

然后我在VENDOR_CODE列上建了一个索引

SQL> create index idx_vendor on sp_trans(vendor_code);

索引已创建。

已用时间: 00: 00: 02.03

已选择8行。

已用时间: 00: 00: 01.42

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=646 Card=1362 Bytes=
128028)
1 0 HASH JOIN (Cost=646 Card=1362 Bytes=128028)
2 1 HASH JOIN (Cost=317 Card=1362 Bytes=76272)
3 2 TABLE ACCESS (BY INDEX ROWID) OF 'SP_TRANS' (Cost=2 Ca
rd=273 Bytes=7098)
4 3 INDEX (RANGE SCAN) OF 'IDX_VENDOR' (NON-UNIQUE) (Cos
t=1 Card=273)
5 2 TABLE ACCESS (FULL) OF 'SP_TRANS_SUB' (Cost=158 Card=1
35502 Bytes=4065060)
6 1 TABLE ACCESS (FULL) OF 'SP_ITEM' (Cost=77 Card=29547 Byt
es=1122786)
Statistics
----------------------------------------------------------
0 recursive calls
8 db block gets
1546 consistent gets
1 physical reads
0 redo size
1732 bytes sent via SQL*Net to client
425 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
3 sorts (memory)
0 sorts (disk)
8 rows processed
SQL>

这时的结果稍好一点，可是不如最初的执行计划。

加rule提示，把后来建的索引删了的结果，这就是最初的执行计划

SQL> SELECT /*+ rule */ "SP_TRANS"."TRANS_NO",
2 "SP_TRANS_SUB"."ITEM_CODE",
3 "SP_ITEM"."ITEM_NAME",
4 "SP_ITEM"."CHART_ID",
5 "SP_ITEM"."SPECIFICATION",
6 "SP_TRANS_SUB"."COUNTRY",
7 "SP_TRANS_SUB"."QTY",
8 "SP_TRANS_SUB"."PRICE",
9 "SP_TRANS"."VENDOR_CODE",
10 "SP_TRANS"."PAY_MODE",
11 NVL("SP_TRANS_SUB"."PAY_QTY",0),
12 0 as PAY_THIS
13 FROM "SP_ITEM",
14 "SP_TRANS_SUB",
15 "SP_TRANS"
16 WHERE ( "SP_TRANS_SUB"."TRANS_NO" = "SP_TRANS"."TRANS_NO" ) and
17 ( "SP_ITEM"."ITEM_CODE" = "SP_TRANS_SUB"."ITEM_CODE" ) and
18 ( ( "SP_TRANS"."VENDOR_CODE" = '20011021023') )
19 /

已选择8行。

已用时间: 00: 00: 00.71

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: RULE
1 0 NESTED LOOPS
2 1 NESTED LOOPS
3 2 TABLE ACCESS (FULL) OF 'SP_TRANS'
4 2 TABLE ACCESS (BY INDEX ROWID) OF 'SP_TRANS_SUB'
5 4 INDEX (RANGE SCAN) OF 'PK_SP_TRANS_SUB' (UNIQUE)
6 1 TABLE ACCESS (BY INDEX ROWID) OF 'SP_ITEM'
7 6 INDEX (UNIQUE SCAN) OF 'PK_SP_ITEM' (UNIQUE)
Statistics
----------------------------------------------------------
0 recursive calls
4 db block gets
323 consistent gets
0 physical reads
0 redo size
1809 bytes sent via SQL*Net to client
426 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
8 rows processed
SQL>

然后创建这个索引，得出的时间大大缩短，看来CBO在某些情况下确实会选择非最优的执行计划

SQL> create index idx_vendor on sp_trans(vendor_code);

索引已创建。

已用时间: 00: 00: 02.43

已选择8行。

已用时间: 00: 00: 00.31

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: RULE
1 0 NESTED LOOPS
2 1 NESTED LOOPS
3 2 TABLE ACCESS (BY INDEX ROWID) OF 'SP_TRANS'
4 3 INDEX (RANGE SCAN) OF 'IDX_VENDOR' (NON-UNIQUE)
5 2 TABLE ACCESS (BY INDEX ROWID) OF 'SP_TRANS_SUB'
6 5 INDEX (RANGE SCAN) OF 'PK_SP_TRANS_SUB' (UNIQUE)
7 1 TABLE ACCESS (BY INDEX ROWID) OF 'SP_ITEM'
8 7 INDEX (UNIQUE SCAN) OF 'PK_SP_ITEM' (UNIQUE)
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
49 consistent gets
1 physical reads
0 redo size
1809 bytes sent via SQL*Net to client
426 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
8 rows processed
SQL>

然而这不是最快的，这是RBO

我们修改optimizer_index_cost_adj参数，然后奇迹出现了

optimizer_index_cost_adj告诉CBO倾向于使用索引扫描而不是全表扫描的程度，该参数缺省值是１００，就是不左右CBO的判断

SQL> alter session set optimizer_index_cost_adj=30
2 /

会话已更改。

已用时间: 00: 00: 00.20

已选择8行。

已用时间: 00: 00: 00.11

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=658 Card=1095 Bytes= 102930)
1 0 NESTED LOOPS (Cost=658 Card=1095 Bytes=102930)
2 1 NESTED LOOPS (Cost=329 Card=1095 Bytes=61320)
3 2 TABLE ACCESS (BY INDEX ROWID) OF 'SP_TRANS' (Cost=1 Card=273 Bytes=7098)
4 3 INDEX (RANGE SCAN) OF 'IDX_VENDOR' (NON-UNIQUE) (Cost=1 Card=273)
5 2 TABLE ACCESS (BY INDEX ROWID) OF 'SP_TRANS_SUB' (Cost= 2 Card=135502 Bytes=4065060)
6 5 INDEX (RANGE SCAN) OF 'PK_SP_TRANS_SUB' (UNIQUE) (Cost=3 Card=135502)
7 1 TABLE ACCESS (BY INDEX ROWID) OF 'SP_ITEM' (Cost=1 Card=29547 Bytes=1122786)
8 7 INDEX (UNIQUE SCAN) OF 'PK_SP_ITEM' (UNIQUE)
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
49 consistent gets
0 physical reads
0 redo size
1809 bytes sent via SQL*Net to client
426 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
8 rows processed
SQL>
SQL> drop index idx_vendor;

索引已丢弃。

已用时间: 00: 00: 00.61

SQL> /

已选择8行。

已用时间: 00: 00: 00.11

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=700 Card=1095 Bytes=
102930)
1 0 NESTED LOOPS (Cost=700 Card=1095 Bytes=102930)
2 1 NESTED LOOPS (Cost=371 Card=1095 Bytes=61320)
3 2 TABLE ACCESS (FULL) OF 'SP_TRANS' (Cost=43 Card=273 By
tes=7098)
4 2 TABLE ACCESS (BY INDEX ROWID) OF 'SP_TRANS_SUB' (Cost=
2 Card=135502 Bytes=4065060)
5 4 INDEX (RANGE SCAN) OF 'PK_SP_TRANS_SUB' (UNIQUE) (Co
st=3 Card=135502)
6 1 TABLE ACCESS (BY INDEX ROWID) OF 'SP_ITEM' (Cost=1 Card=
29547 Bytes=1122786)
7 6 INDEX (UNIQUE SCAN) OF 'PK_SP_ITEM' (UNIQUE)
Statistics
----------------------------------------------------------
0 recursive calls
4 db block gets
323 consistent gets
0 physical reads
0 redo size
1809 bytes sent via SQL*Net to client
426 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
8 rows processed
SQL>

关于optimizer_index_cost_adj等影响CBO的参数及设置，richto介绍过一篇好文章：

http://www.evdbt.com/SearchIntelligenceCBO.doc

关于成本的计算等，请参考以下文章：

http://www.centrexcc.com/A%20Loo ... 2010053%20Event.pdf
http://www.centrexcc.com/A%20Loo ... 2010053%20Event.ppt

很简单的问题，希望对大家有点帮助!