latch: cache buffers chains故障处理总结

客户说数据库的CPU使用率为100%，应用相应迟缓。
发现是latch: cache buffers chains 作祟
故障分析思路

查看等待事件，判断故障起因

1 SQL>select * from (select sid,event,p1,p2,p3,p1text,WAIT_TIME,SECONDS_IN_WAIT from v$session_wait where wait_class# <> 6
2 order by wait_time desc) where rownum <=10;

确认为latch: cache buffers chains引起的故障后，查看latch的命中率

1 SQL>SELECT name, gets, misses, sleeps,
2 immediate_gets, immediate_misses
3 FROM v$latch
4 WHERE name = 'cache buffers chains';

各列名称意义如下

1 NAME：latch名称
2 IMMEDIATE_GETS：以Immediate模式latch请求数
3 IMMEDIATE_MISSES：请求失败数
4 GETS：以Willing to wait请求模式latch的请求数
5 MISSES：初次尝试请求不成功次数
6 SPIN_GETS：第一次尝试失败，但在以后的轮次中成功
7 SLEEP[x]：成功获取前sleeping次数
8 WAIT_TIME：花费在等待latch的时间

这里需要注意MISSES/GETS如果在达10%左右，则说明有比较严重的latch争用,也可以通过查询v$latch_children视图查看其他latch信息 ，语句如下

1 SQL> SELECT *
2 FROM (SELECT addr, child#, gets, misses, sleeps, immediate_gets igets,
3 immediate_misses imiss, spin_gets sgets
4 FROM v$latch_children
5 WHERE NAME = 'cache buffers chains'
6 ORDER BY sleeps DESC)
7 WHERE ROWNUM < 11;

关于latch的统计信息，主要关注以下几部分
misses/gets的比率是多少
获自spinning的misses的百分比是多少
latch请求了多少次
latch休眠了多少次

查看热点对象和访问信息，TCH列表示对象被访问的次数

 1 SQL> SELECT *
 2 FROM ( SELECT addr,
 3 ts#,
 4 file#,
 5 dbarfil,
 6 dbablk,
 7 tch
 8 FROM x$bh
 9 ORDER BY tch DESC)
10 WHERE ROWNUM < 11;

通过对象的文件号和块号查看具体对象信息

1 SQL>select owner, segment_name, partition_name, tablespace_name
2 from dba_extents
3 where relative_fno = &v_dba_rfile
4 and &v_dba_block between block_id and block_id + blocks - 1;

也可以通过如下sql查找热点块，主要

1 SELECT *
2 FROM (SELECT O.OWNER, O.OBJECT_NAME, O.OBJECT_TYPE, SUM(TCH) TOUCHTIME
3 FROM X$BH B, DBA_OBJECTS O
4 WHERE B.OBJ = O.DATA_OBJECT_ID
5 AND B.TS# > 0
6 GROUP BY O.OWNER, O.OBJECT_NAME, O.OBJECT_TYPE
7 ORDER BY SUM(TCH) DESC)
8 WHERE ROWNUM <= 10;

查看引起latch: cache buffers chains的sql

 1 SQL> select * from (select
 2 count(*),
 3 sql_id,
 4 nvl(o.object_name,ash.current_obj#) objn,
 5 substr(o.object_type,0,10) otype,
 6 CURRENT_FILE# fn,
 7 CURRENT_BLOCK# blockn
 8 from v$active_session_history ash
 9 , all_objects o
10 where event like 'latch: cache buffers chains'
11 and o.object_id (+)= ash.CURRENT_OBJ#
12 group by sql_id, current_obj#, current_file#,
13 current_block#, o.object_name,o.object_type
14 order by count(*) desc )where rownum <=10;

根据上面得到的sql_id信息查看sql全文

1 SQL>select sql_fulltext from v$sqlarea where sql_id='&sqlid';

查看SQL的执行计划

1 SQL>SELECT * FROM table(DBMS_XPLAN.DISPLAY_CURSOR(('&sql_id',0));

在认为sql执行计划不准确的情况也可以通过sql_id查看sql的address和hash_value查看sql的实际执行计划

1 SQL>SELECT address, hash_value FROM v$sql
2 WHERE sql_id='&sql_id';
3 SQL>SELECT operation, options, object_name, cost FROM v$sql_plan
4 WHERE address = '&addr' AND hash_value = 'hash_v';

当某个会话长时间持有latch时，可以通过联合v$latchholder和v$session视图查看sql信息

1 SQL>SELECT s.sql_hash_value,s.sql_id,s.address, l.name
2 FROM V$SESSION s, V$LATCHHOLDER l
3 WHERE s.sid = l.sid;

故障处理思路

1、根据sql执行计划判断该执行计划是否正确，sql执行过长往往意味着过长时间的持有latch。
2、优化nested loop join，如果有可能使用hash join代替nested loop join。也可以利用对热块索引进行hash分区，或者使用hash簇的方式减缓热块现象。
3、调整表的pctfree值，将数据尽可能的分布到多个块中
4、调整应用