最近忙着把公司的数据库从mysql迁移到oracle,期间作了很多工作来优化oracle平台的性能,不过这里面最大的性能调整还是来自sql。下面举一个web翻页sql调整的例子。
环境:
Linux version 2.4.20-8custom (root@web2) (gcc version 3.2.2 20030222 (Red Hat Linux 3.2.2-5)) #3 SMP Thu Jun 5 22:03:36 CST 2003
Mem: 2113466368
Swap: 4194881536
CPU:两个超线程的Intel(R) Xeon(TM) CPU 2.40GHz
优化前语句在mysql里面查询15秒左右出来,转移到oracle后进行在不调整索引和语句的情况下执行时间大概是4-5秒,调整后执行时间小于0.5秒。
翻页语句:
SELECT * FROM (SELECT T1.*, rownum as linenum FROM (
SELECT /*+ index(a ind_old)*/
a.category FROM auction_auctions a WHERE a.category =' 170101 ' AND a.closed='0' AND ends > sysdate AND (a.approve_status>=0) ORDER BY a.ends) T1 WHERE rownum < 18681) WHERE linenum >= 18641
被查询的表:auction_auctions(产品表)
表结构:
Code: | [Copy to clipboard] | |
|
表记录数及大小
SQL> select count(*) from auction_auctions;
COUNT(*)
----------
537351
SQL> select segment_name,bytes,blocks from user_segments where segment_name ='AUCTION_AUCTIONS';
SEGMENT_NAME BYTES BLOCKS
AUCTION_AUCTIONS 1059061760 129280
表上原有的索引
create index ind_old on auction_auctions(closed,approve_status,category,ends) tablespace tbsindex compress 2;
SQL> select segment_name,bytes,blocks from user_segments where segment_name = 'IND_OLD';
SEGMENT_NAME BYTES BLOCKS
IND_OLD 20971520 2560
表和索引都已经分析过,我们来看一下sql执行的费用
SQL> set autotrace trace;
SQL> SELECT * FROM (SELECT T1.*, rownum as linenum FROM (SELECT a.* FROM auction_auctions a WHERE a.category like '18%' AND a.closed='0' AND ends > sysdate AND (a.approve_status>=0) ORDER BY a.ends) T1 WHERE rownum <18681) WHERE linenum >= 18641;
40 rows selected.
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=19152 Card=18347 Byt
es=190698718)
1 0 VIEW (Cost=19152 Card=18347 Bytes=190698718)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=19152 Card=18347 Bytes=190460207)
4 3 TABLE ACCESS (BY INDEX ROWID) OF 'AUCTION_AUCTIONS'
(Cost=19152 Card=18347 Bytes=20860539)
5 4 INDEX (RANGE SCAN) OF 'IND_OLD' (NON-UNIQUE) (Cost
=810 Card=186003)
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
19437 consistent gets
18262 physical reads
0 redo size
114300 bytes sent via SQL*Net to client
56356 bytes received via SQL*Net from client
435 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
40 rows processed
我们可以看到这条sql语句通过索引范围扫描找到最里面的结果集,然后通过两个view操作最后得出数据。其中18502 consistent gets,17901 physical reads
我们来看一下这个索引建的到底合不合理,先看下各个查寻列的distinct值
select count(distinct ends) from auction_auctions;
COUNT(DISTINCTENDS)
-------------------
338965
SQL> select count(distinct category) from auction_auctions;
COUNT(DISTINCTCATEGORY)
-----------------------
1148
SQL> select count(distinct closed) from auction_auctions;
COUNT(DISTINCTCLOSED)
---------------------
2
SQL> select count(distinct approve_status) from auction_auctions;
COUNT(DISTINCTAPPROVE_STATUS)
-----------------------------
5
页索引里列平均存储长度
SQL> select avg(vsize(ends)) from auction_auctions;
AVG(VSIZE(ENDS))
----------------
7
SQL> select avg(vsize(closed)) from auction_auctions;
AVG(VSIZE(CLOSED))
------------------
2
SQL> select avg(vsize(category)) from auction_auctions;
AVG(VSIZE(CATEGORY))
--------------------
5.52313106
SQL> select avg(vsize(approve_status)) from auction_auctions;
AVG(VSIZE(APPROVE_STATUS))
--------------------------
1.67639401
我们来估算一下各种组合索引的大小,可以看到closed,approve_status,category都是相对较低集势的列(重复值较多),下面我们来大概计算下各种页索引需要的空间
column distinct num column len
ends 338965 7
category 1148 5.5
closed 2 2
approve_status 5 1.7
index1: (ends,closed,category,approve_status) compress 2
ends:distinct number---338965
closed: distinct number---2
index size=338965*2*(9+2)+ 537351*(1.7+5.5+6)=14603998
index2: (closed,category,ends,approve_status)
closed: distinct number---2
category: distinct number---1148
index size=2*1148*(2+5.5)+537351*(7+1.7+6)=7916279
index3: (closed,approve_status,category,ends)
closed: distinct number---2
approve_status: distinct number—5
index size=2*5*(2+1.7)+537351*(7+5.5+6)=9941030
结果出来了,index2: (closed,category,ends,approve_status)的索引最小
我们再来看一下语句
SELECT * FROM (SELECT T1.*, rownum as linenum FROM (SELECT a.* FROM auction_auctions a WHERE a.category like '18%' AND a.closed='0' AND ends > sysdate AND (a.approve_status>=0) ORDER BY a.ends) T1 WHERE rownum <18681) WHERE linenum >= 18641;
可以看出这个sql语句有很大优化余地,首先最里面的结果集SELECT a.* FROM auction_auctions a WHERE a.category like '18%' AND a.closed='0' AND ends > sysdate AND (a.approve_status>=0) ORDER BY a.ends,这里的话会走index range scan,然后table scan by rowid,这样的话如果符合条件的数据多的话相当耗资源,我们可以改写成
SELECT a.rowid FROM auction_auctions a WHERE a.category like '18%' AND a.closed='0' AND ends > sysdate AND (a.approve_status>=0) ORDER BY a.ends
这样的话最里面的结果集只需要index fast full scan就可以完成了,再改写一下得出以下语句
select * from auction_auctions where rowid in (SELECT rid FROM (
SELECT T1.rowid rid, rownum as linenum FROM
(SELECT a.rowid FROM auction_auctions a WHERE a.category like '18%' AND a.closed='0' AND ends > sysdate AND
(a.approve_status>=0) ORDER BY a.ends) T1 WHERE rownum < 18681) WHERE linenum >= 18641)
下面我们来测试一下这个索引的查询开销
select * from auction_auctions where rowid in (SELECT rid FROM (
SELECT T1.rowid rid, rownum as linenum FROM
(SELECT a.rowid FROM auction_auctions a WHERE a.category like '18%' AND a.closed='0' AND ends > sysdate AND
(a.approve_status>=0) ORDER BY a.closed,a.ends) T1 WHERE rownum < 18681) WHERE linenum >= 18641)
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=18698 Card=18344 Byt
es=21224008)
1 0 NESTED LOOPS (Cost=18698 Card=18344 Bytes=21224008)
2 1 VIEW (Cost=264 Card=18344 Bytes=366880)
3 2 SORT (UNIQUE)
4 3 COUNT (STOPKEY)
5 4 VIEW (Cost=264 Card=18344 Bytes=128408)
6 5 SORT (ORDER BY STOPKEY) (Cost=264 Card=18344 Byt
es=440256)
7 6 INDEX (FAST FULL SCAN) OF 'IDX_AUCTION_BROWSE'
(NON-UNIQUE) (Cost=159 Card=18344 Bytes=440256)
8 1 TABLE ACCESS (BY USER ROWID) OF 'AUCTION_AUCTIONS' (Cost
=1 Card=1 Bytes=1137)
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
2080 consistent gets
1516 physical reads
0 redo size
114840 bytes sent via SQL*Net to client
56779 bytes received via SQL*Net from client
438 SQL*Net roundtrips to/from client
2 sorts (memory)
0 sorts (disk)
40 rows processed
可以看到consistent gets从19437降到2080,physical reads从18262降到1516,查询时间也丛4秒左右下降到0。5秒,可以来说这次sql调整取得了预期的效果。
今天又修改了一下语句,
SQL> select * from auction_auctions where rowid in
2 (SELECT rid FROM (
3 SELECT T1.rowid rid, rownum as linenum FROM
4 (SELECT a.rowid FROM auction_auctions a
5 WHERE a.category like '18%' AND a.closed='0' AND ends > sysdate AND
a.approve_status>=0
6 7 ORDER BY a.closed,a.category,a.ends) T1
8 WHERE rownum < 18600) WHERE linenum >= 18560) ;
40 rows selected.
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=17912 Card=17604 Byt
es=20367828)
1 0 NESTED LOOPS (Cost=17912 Card=17604 Bytes=20367828)
2 1 VIEW (Cost=221 Card=17604 Bytes=352080)
3 2 SORT (UNIQUE)
4 3 COUNT (STOPKEY)
5 4 VIEW (Cost=221 Card=17604 Bytes=123228)
6 5 INDEX (RANGE SCAN) OF 'IDX_AUCTION_BROWSE' (NON-
UNIQUE) (Cost=221 Card=17604 Bytes=422496)
7 1 TABLE ACCESS (BY USER ROWID) OF 'AUCTION_AUCTIONS' (Cost
=1 Card=1 Bytes=1137)
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
550 consistent gets
14 physical reads
0 redo size
117106 bytes sent via SQL*Net to client
56497 bytes received via SQL*Net from client
436 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
40 rows processed
在order by里加上索引前导列,消除了
6 5 SORT (ORDER BY STOPKEY) (Cost=264 Card=18344 Byt
es=440256)
,把consistent gets从2080降到550