环境:oracle 817 + + 阵列柜
swd_billdetail 表5000万条数据
SUPER_USER 表2800条数据
连接列上都有索引,而且super_user中的一条对应于swd_billdetail表中的很多条记录
表与索引都做了分析。
实际应用的查询为:
select a.CHANNEL, B.user_class
from swd_billdetail B, SUPER_USER A
where A.cn = B.cn;
这样在分析时导致查询出的数据过多,不方便,所以用count(a.CHANNEL||B.user_class)来代替,而且count(a.CHANNEL||B.user_class)操作本身并不占用过多的时间,所以可以接受此种替代。
利用索引查询出SWD_BILLDETAIL表中所有记录的方法
SQL> select count(id) from SWD_BILLDETAIL;
COUNT(ID)
----------
53923574
Elapsed: 00:02:166.00
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=18051 Card=1)
1 0 SORT (AGGREGATE)
2 1 INDEX (FAST FULL SCAN) OF 'SYS_C001851' (UNIQUE) (Cost=18051 Card=54863946)
Statistics
----------------------------------------------------------
0 recursive calls
1952 db block gets
158776 consistent gets
158779 physical reads
1004 redo size
295 bytes sent via SQL*Net to client
421 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
1 rows processed
利用全表扫描从SWD_BILLDETAIL表中取出全部数据的方法。
SQL> select count(user_class) from swd_billdetail;
COUNT(USER_CLASS)
-----------------
53923574
Elapsed: 00:11:703.07
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=165412 Card=1 Bytes=2)
1 0 SORT (AGGREGATE)
2 1 TABLE ACCESS (FULL) OF 'SWD_BILLDETAIL' (Cost=165412 Card=54863946 Bytes=109727892)
Statistics
----------------------------------------------------------
0 recursive calls
8823 db block gets
1431070 consistent gets
1419520 physical reads
0 redo size
303 bytes sent via SQL*Net to client
421 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
1 rows processed
select count(a.CHANNEL||B.user_class)
from swd_billdetail B, SUPER_USER A
where A.cn = B.cn;
EXEC_ORDER PLANLINE
---------- -----------------------------------------------------------------------------------------------------------
6 SELECT STATEMENT OPT_MODE:CHOOSE (COST=108968,CARD=1,BYTES=21)
5 SORT (AGGREGATE) (COST=,CARD=1,BYTES=21)
4 NESTED LOOPS (COST=108968,CARD=1213745,BYTES=25488645)
1 TABLE ACCESS (FULL) OF 'SWORD.SUPER_USER' (COST=2,CARD=2794,BYTES=27940)
3 TABLE ACCESS (BY INDEX ROWID) OF 'SWORD.SWD_BILLDETAIL' (COST=39,CARD=54863946,BYTES=603503406)
2 INDEX (RANGE SCAN) OF 'SWORD.IDX_DETAIL_CN' (NON-UNIQUE) (COST=3,CARD=54863946,BYTES=)
这个查询耗费的时间很长,需要1个多小时。
运行后的信息如下:
COUNT(A.CHANNEL||B.USER_CLASS)
------------------------------
1186387
Elapsed: 01:107:6429.87
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=108968 Card=1 Bytes=21)
1 0 SORT (AGGREGATE)
2 1 NESTED LOOPS (Cost=108968 Card=1213745 Bytes=25488645)
3 2 TABLE ACCESS (FULL) OF 'SUPER_USER' (Cost=2 Card=2794Bytes=27940)
4 2 TABLE ACCESS (BY INDEX ROWID) OF 'SWD_BILLDETAIL' (Cost=39 Card=54863946 Bytes=603503406)
5 4 INDEX (RANGE SCAN) OF 'IDX_DETAIL_CN' (NON-UNIQUE) (Cost=3 Card=54863946)
Statistics
----------------------------------------------------------
0 recursive calls
4 db block gets
1196954 consistent gets
1165726 physical reads
0 redo size
316 bytes sent via SQL*Net to client
421 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
2 sorts (memory)
0 sorts (disk)
1 rows processed
将语句中加入hints,让oracle的优化器使用嵌套循环,并且大表作为驱动表,生成新的执行计划:
select /*+ ORDERED USE_NL(A) */ count(a.CHANNEL||B.user_class)
from swd_billdetail B, SUPER_USER A
where A.cn = B.cn;
EXEC_ORDER PLANLINE
---------- -----------------------------------------------------------------------------------------------------
6 SELECT STATEMENT OPT_MODE:CHOOSE (COST=109893304,CARD=1,BYTES=21)
5 SORT (AGGREGATE) (COST=,CARD=1,BYTES=21)
4 NESTED LOOPS (COST=109893304,CARD=1213745,BYTES=25488645)
1 TABLE ACCESS (FULL) OF 'SWORD.SWD_BILLDETAIL' (COST=165412,CARD=54863946,BYTES=603503406)
3 TABLE ACCESS (BY INDEX ROWID) OF 'SWORD.SUPER_USER' (COST=2,CARD=2794,BYTES=27940)
2 INDEX (RANGE SCAN) OF 'SWORD.IDX_SUPER_USER_CN' (NON-UNIQUE) (COST=1,CARD=2794,BYTES=)
这个查询耗费的时间较短,才20分钟,性能比较好。运行后的信息如下:
COUNT(A.CHANNEL||B.USER_CLASS)
------------------------------
1186387
Elapsed: 00:20:1208.87
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=109893304 Card=1 Bytes=21)
1 0 SORT (AGGREGATE)
2 1 NESTED LOOPS (Cost=109893304 Card=1213745 Bytes=25488645)
3 2 TABLE ACCESS (FULL) OF 'SWD_BILLDETAIL' (Cost=165412 Card=54863946 Bytes=603503406)
4 2 TABLE ACCESS (BY INDEX ROWID) OF 'SUPER_USER' (Cost=2Card=2794 Bytes=27940)
5 4 INDEX (RANGE SCAN) OF 'IDX_SUPER_USER_CN' (NON-UNIQUE) (Cost=1 Card=2794)
Statistics
----------------------------------------------------------
0 recursive calls
8823 db block gets
56650250 consistent gets
1413250 physical reads
0 redo size
316 bytes sent via SQL*Net to client
421 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
2 sorts (memory)
0 sorts (disk)
1 rows processed
总结: 因为上两个查询都是采用nested loop循环,这时采用哪个表作为driving table就很重要。在第一个sql中,小表(SUPER_USER)作为driving table,符合oracle优化的建议,但是由于SWD_BILLDETAIL表中cn列的值有很多重复的,这样对于SUPER_USER中的每一行,都会在SWD_BILLDETAIL中有很多行,利用索引查询出这些行的rowid很快,但是再利用这些rowid去查询SWD_BILLDETAIL表中的user_class列的值,就比较慢了。原因是这些rowid是随机的,而且该表比较大,不可能缓存到内存,所以几乎每次按照rowid查询都需要读物理磁盘,这就是该执行计划比较慢的真正原因。从结果可以得到验证:查询出1186387行,需要利用rowid从SWD_BILLDETAIL表中读取1186387次,而且大部分为从硬盘上读取。
反其道而行之,利用大表(SWD_BILLDETAIL)作为driving表,这样大表只需要做一次全表扫描(而且会使用多块读功能,每次物理I/O都会读取几个oracle数据块,从而一次读取很多行,加快了执行效率),对于读出的每一行,都与SUPER_USER中的行进行匹配,因为SUPER_USER表很小,所以可以全部放到内存中,这样匹配操作就极快,所以该sql执行的时间与SWD_BILLDETAIL表全表扫描的时间差不多(SWD_BILLDETAIL全表用11分钟,而此查询用20分钟)。
另外:如果SWD_BILLDETAIL表中cn列的值唯一,则第一个sql执行计划执行的结果或许也会不错。如果SUPER_USER表也很大,如500万行,则第2个sql执行计划执行的结果反而又可能会差。其实,如果SUPER_USER表很小,则第2个sql语句的执行计划如果不利用SUPER_USER表的索引,查询或许会更快一些,我没有对此进行。
所以在进行性能调整时,具体问题要具体分析,没有一个统一的标准。
【责编:yuan】
--------------------next---------------------