这篇论坛文章(赛迪网技术社区)针对高级复制中由于mlog$表的HWM过高导致的性能问题进行了详细的讲解,更多内容请参考下文:
某系统升级后,报告事务端的数据库反应非常慢,用户大量投诉。
现场检查系统后,发现CPU占用到了80%(平时该系统正常时只有10%~20%)。wio也非常高(说明磁盘读写严重)。一段时间后,现场将statspack报告发回来了:
STATSPACK report for
DB Name DB Id Instance Inst Num Release Cluster Host
------------ ----------- ------------ -------- ----------- ------- ------------
OLTPDB 3781951398 OLTPDB 1 9.2.0.5.0 NO jnaip21
Snap Id Snap Time Sessions Curs/Sess Comment
------- ------------------ -------- --------- -------------------
Begin Snap: 3 27-Sep-04 13:15:58 32 83.4
End Snap: 4 27-Sep-04 14:10:36 37 83.4
Elapsed: 54.63 (mins)
Cache Sizes (end)
~~~~~~~~~~~~~~~~~
Buffer Cache: 720M Std Block Size: 8K
Shared Pool Size: 112M Log Buffer: 512K
Load Profile
~~~~~~~~~~~~ Per Second Per Transaction
--------------- ---------------
Redo size: 11,763.97 7,400.17
Logical reads: 8,251.68 5,190.75
Block changes: 83.17 52.32
Physical reads: 2,488.23 1,565.23
Physical writes: 6.15 3.87
User calls: 411.88 259.09
Parses: 56.57 35.59
Hard parses: 0.29 0.18
Sorts: 14.59 9.18
Logons: 0.12 0.08
Executes: 80.04 50.35
Transactions: 1.59
% Blocks changed per Read: 1.01 Recursive Call %: 28.14
Rollback per transaction %: 0.17 Rows per Sort: 1818.01
Instance Efficiency Percentages (Target 100%)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Buffer Nowait %: 99.99 Redo NoWait %: 100.00
Buffer Hit %: 69.88 In-memory Sort %: 99.99
Library Hit %: 99.78 Soft Parse %: 99.50
Execute to Parse %: 29.32 Latch Hit %: 99.98
Parse CPU to Parse Elapsd %: 96.72 % Non-Parse CPU: 99.02
Shared Pool Statistics Begin End
------ ------
Memory Usage %: 93.60 93.71
% SQL with executions>1: 20.28 19.05
% Memory for SQL w/exec>1: 25.59 25.88
Top 5 Timed Events
~~~~~~~~~~~~~~~~~~ % Total
Event Waits Time (s) Ela Time
-------------------------------------------- ------------ ----------- --------
db file scattered read 599,637 3,886 63.79
CPU time 1,142 18.74
db file sequential read 238,983 629 10.33
enqueue 263 366 6.00
log file sync 5,154 23 .37
-------------------------------------------------------------
Wait Events for DB: OLTPDB Instance: OLTPDB Snaps: 3 -4
-> s - second
-> cs - centisecond - 100th of a second
-> ms - millisecond - 1000th of a second
-> us - microsecond - 1000000th of a second
-> ordered by wait time desc, waits desc (idle events last)
Avg
Total Wait wait Waits
Event Waits Timeouts Time (s) (ms) /txn
---------------------------- ------------ ---------- ---------- ------ --------
db file scattered read 599,637 0 3,886 6 115.1
db file sequential read 238,983 0 629 3 45.9
enqueue 263 4 366 1390 0.1
log file sync 5,154 0 23 4 1.0
log file parallel write 15,143 8,060 20 1 2.9
db file parallel write 1,008 0 12 12 0.2
control file parallel write 1,065 0 4 4 0.2
buffer busy waits 3,195 0 4 1 0.6
... ...
SQL ordered by Gets for DB: OLTPDB Instance: OLTPDB Snaps: 3 -4
-> End Buffer Gets Threshold: 10000
-> Note that resources reported for PL/SQL includes the resources used by
all SQL statements called within the PL/SQL code. As individual SQL
statements are also reported, it is possible and valid for the summed
total % to exceed 100
CPU Elapsd
Buffer Gets Executions Gets per Exec %Total Time (s) Time (s) Hash Value
--------------- ------------ -------------- ------ -------- --------- ----------
6,509,469 1,060 6,141.0 24.1 66.35 64.61 2833840857
Module: JDBC Thin Client
select lib.code, lib.name, lib.info, leaf.leafdirindex, tp.tpname, from t_commend com, t_leafinfo leaf, t_lib lib,t_tpinfo tp where
2,258,385 343 6,584.2 8.3 65.41 1289.82 3307794804
Module: oracle@qdaip12 (TNS V1-V3)
SELECT /*+ */ "A2"."PHONENUMBER","A2"."TGROUPID","A2"."LOOPNO
","A2"."TCODE" FROM "RU"."T_LOOP" "A2", (SELECT /*+ */
DISTINCT "A3"."LOOPNO" "LOOPNO","A3"."TGROUPID" "TGROUPID"
FROM "RU"."MLOG$_T_LOOP" "A3" WHERE "A3"."SNAPTIME$$">:1
AND "A3"."DMLTYPE$$"<>'D') "A1" WHERE "A2"."LOOPNO"="A1"."LOOPNO
2,251,917 343 6,565.4 8.3 66.51 1306.09 1606105459
Module: oracle@qdaip12 (TNS V1-V3)
SELECT /*+ */ DISTINCT "A1"."LOOPNO","A1"."TGROUPID" FROM "RI
NG"."MLOG$_T_LOOP" "A1" WHERE (("A1"."LOOPNO","A1"."TGROU
PID")<>ALL (SELECT "A2"."LOOPNO","A2"."TGROUPID" FROM "RU".
"T_LOOP" "A2" WHERE "A1"."LOOPNO"="A2"."LOOPNO" AND "A1"."TO
NEGROUPID"="A2"."TGROUPID")) AND "A1"."SNAPTIME$$">:1 AND "A1
2,249,895 343 6,559.5 8.3 66.34 1098.20 4039622144
delete from "RU"."MLOG$_T_LOOP" where snaptime$$ <= :1
... ...
Physical Reads Executions Reads per Exec %Total Time (s) Time (s) Hash Value
--------------- ------------ -------------- ------ -------- --------- ----------
2,129,843 343 6,209.5 26.1 66.34 1098.20 4039622144
delete from "RU"."MLOG$_T_LOOP" where snaptime$$ <= :1
2,129,004 343 6,207.0 26.1 65.41 1289.82 3307794804
Module: oracle@qdaip12 (TNS V1-V3)
SELECT /*+ */ "A2"."PHONENUMBER","A2"."TGROUPID","A2"."LOOPNO
","A2"."TCODE" FROM "RU"."T_LOOP" "A2", (SELECT /*+ */
DISTINCT "A3"."LOOPNO" "LOOPNO","A3"."TGROUPID" "TGROUPID"
FROM "RU"."MLOG$_T_LOOP" "A3" WHERE "A3"."SNAPTIME$$">:1
AND "A3"."DMLTYPE$$"<>'D') "A1" WHERE "A2"."LOOPNO"="A1"."LOOPNO
2,123,001 343 6,189.5 26.0 66.51 1306.09 1606105459
Module: oracle@qdaip12 (TNS V1-V3)
SELECT /*+ */ DISTINCT "A1"."LOOPNO","A1"."TGROUPID" FROM "RI
NG"."MLOG$_T_LOOP" "A1" WHERE (("A1"."LOOPNO","A1"."TGROU
PID")<>ALL (SELECT "A2"."LOOPNO","A2"."TGROUPID" FROM "RU".
"T_LOOP" "A2" WHERE "A1"."LOOPNO"="A2"."LOOPNO" AND "A1"."TO
NEGROUPID"="A2"."TGROUPID")) AND "A1"."SNAPTIME$$">:1 AND "A1
1,724,081 278 6,201.7 21.1 54.05 887.05 2153620124
update "RU"."MLOG$_T_LOOP" set snaptime$$ = :1 where snap
time$$ > to_date('2100-01-01:00:00:00','YYYY-MM-DD:HH24:MI:SS')
关注statspack中以下信息:Buffer Hit非常低,只有69.88%;在Top Event的等待事件中,db file scattered read的等待最高,占到了63.79%。很明显,这是由于数据文件扫描导致的buffer cache下降非常严重。而就肯定是由于存在全表扫描(全表扫描到的数据块不会被cache在buffer cache,而是会已最快的速度被置换出内存,这就是为什么全表扫描会导致buffer hit下降)语句所致。
再看top sql中的内容,我们发现,大多数top sql都涉及到一个奇特的表:MLOG$_T_LOOP。这张表的使用是不会出现在应用程序中。因此相关语句都是被系统调用的。那么这张表是做什么用的呢?
我们的系统为了将业务处理与业务查询分离,通过高级复制将业务处理端的某些表同步到业务查询端去。而mlog$_表就是记录最后一次刷新后,被同步表的数据变化的日志表,当查询端(即快照端)的刷新作业(我们设置10秒钟间隔运行)通过mlog$_读取到源表的数据变化,更新查询端相应的表。而业务处理端检测到所有快照端都更新过这些数据后,就会将这些数据删除。此外,mlog表上是没有任何索引,因此对mlog表的查询肯定是全表扫描。但是由于我们的刷新作业10秒钟更新一次,而10秒钟所产生的数据变化量非常小,在正确情况下,这样的全表扫描是不会影响到系统性能的。
我查了一下MLOG$_T_LOOP,发现它记录的数据确实非常少(一般只有几条数据),但是对它查询钟,physical reads却达到了6200多,这显然不正常。说明这张表的高水位(HWM,高水位记录的是表曾经达到的最大数据块,就像洪水过后留下的水位线。而对表的扫描并不是以当前的实际水位为基准的,而是以高水位为基准的,也就是说会扫描表曾经占用过的所有数据块)非常高。
通过现场了解,我们得知现场在做升级时,并没有停止同步刷新作业,而是将刷新作业间隔延长到2个小时。升级过程中,两边的数据库都没有停止。但是,期间在业务处理端做了一些数据处理,特别是表T_LOOP有大量的数据更新。
问题很明确了!由于升级过程中T_LOOP有大量的数据更新,而快照没有被刷新。因此mlog$_t_loop中保存了两个小时内的所有数据变化记录,导致高水位上升!而刷新作业再次恢复到10秒间隔后,尽管mlog$_t_loop中的数据量下降了,但是它的高水位并没有下降!
于是安排现场将业务处理端的客户连接断开,然后停止快照端的刷新作业,手工刷新一次同步后,检查业务处理端的mlog$表,确保这些表为空,然后truncate mlog$表,恢复同步作业,最后恢复业务处理端的客户连接。经过一段时间观察后,数据库恢复正常。