在很多情况下,当数据库发生性能问题的时候,我们并没有机会来收集足够的诊断信息,比如system state dump或者hang analyze,甚至问题发生的时候DBA根本不在场。这给我们诊断问题带来很大的困难。那么在这种情况下,我们是否能在事后收集一些信息来分析问题的原因呢?
在Oracle 10G或者更高版本上,答案是肯定的。本文我们将介绍一种通过dba_hist_active_sess_history的数据来分析问题的一种方法。
Oracle 10G或更高版本的单机或rac,本文适用于任何平台。
Active Session: on the CPU or was waiting for an event that didn't belong to the Idle wait class
在Oracle 10G中,我们引入了AWR和ASH采样机制,有一个视图gv$active_session_history会每秒钟将数据库所有节点的Active Session采样一次,
而dba_hist_active_sess_history则会将gv$active_session_history里的数据每10秒采样一次并持久化保存。
基于这个特征,我们可以通过分析dba_hist_active_sess_history的Session采样情况,来定位问题发生的准确时间范围,并且可以观察每个采样点的top event和top holder。下面通过一个例子来详细说明。
为了不影响生产系统,我们可以将问题大概期间的ASH数据export出来在测试机上分析。
基于dba_hist_active_sess_history创建一个新表t_ash,然后将其通过exp/imp导入到测试机。在发生问题的数据库上执行exp:
SQL> conn user/passwd
SQL> create table t_ash as select * from dba_hist_active_sess_history where SAMPLE_TIME between TO_TIMESTAMP ('<time_begin>', 'YYYY-MM-DD HH24:MI:SS') and TO_TIMESTAMP ('<time_end>', 'YYYY-MM-DD HH24:MI:SS');
$ exp user/passwd file=t_ash.dmp tables=(t_ash) log=t_ash.exp.log
然后导入到测试机:
$ imp user/passwd file=t_ash.dmp log=t_ash.imp.log
建议采用Oracle SQL Developer来查询以防止输出结果折行不便于观察。
set line 200 pages 1000
col sample_time for a25
col event for a40
alter session set nls_timestamp_format='yyyy-mm-dd hh24:mi:ss.ff';
select
t.dbid, t.instance_number, min(sample_time), max(sample_time), count(*) session_count
from t_ash t
group by t.dbid, t.instance_number
order by dbid, instance_number;
INSTANCE_NUMBER MIN(SAMPLE_TIME) MAX(SAMPLE_TIME) SESSION_COUNT
1 2015-03-26 21:00:04.278 2015-03-26 22:59:48.387 2171
2 2015-03-26 21:02:12.047 2015-03-26 22:59:42.584 36
从以上输出可知该数据库共2个节点,采样时间共2小时,节点1的采样比节点2要多很多,问题可能发生在节点1上。
参考如下脚本:
col event for a35
set lines 222
set pages 9999
select
dbid, instance_number, sample_id, sample_time, count(*) session_count
from t_ash t
group by dbid, instance_number, sample_id, sample_time
order by dbid, instance_number, sample_time;
INSTANCE_NUMBER SAMPLE_ID SAMPLE_TIME SESSION_COUNT
1 36402900 2015-03-26 22:02:50.985 4
1 36402910 2015-03-26 22:03:01.095 1
1 36402920 2015-03-26 22:03:11.195 1
1 36402930 2015-03-26 22:03:21.966 21
1 36402940 2015-03-26 22:03:32.116 102
1 36402950 2015-03-26 22:03:42.226 181
1 36402960 2015-03-26 22:03:52.326 200
1 36402970 2015-03-26 22:04:02.446 227
1 36402980 2015-03-26 22:04:12.566 242
1 36402990 2015-03-26 22:04:22.666 259
1 36403000 2015-03-26 22:04:32.846 289
1 36403010 2015-03-26 22:04:42.966 147
1 36403020 2015-03-26 22:04:53.076 2
1 36403030 2015-03-26 22:05:03.186 4
1 36403040 2015-03-26 22:05:13.296 1
1 36403050 2015-03-26 22:05:23.398 1
注意观察以上输出的每个采样点的active session的数量,数量突然变多往往意味着问题发生了。从以上输出可以确定问题发生的精确时间在2015-03-26 22:03:21 ~ 22:04:42,问题持续了大约1.5分钟。
注意: 观察以上的输出有无断档,比如某些时间没有采样。
在这里我们指定的是top 2 event 并且注掉了采样时间以观察所有采样点的情况。如果数据量较多,您也可以通过开启sample_time的注释来观察某个时间段的情况。注意最后一列session_count指的是该采样点上的等待该event的session数量。
col event for a35
set lines 1000
set pages 9999
select t.dbid,
t.sample_id,
t.sample_time,
t.instance_number,
t.event,
t.session_state,
t.c session_count
from (select t.*,
rank() over(partition by dbid, instance_number, sample_time order by c desc) r
from (select /*+ parallel 2 */
t.*,
count(*) over(partition by dbid, instance_number, sample_time, event) c,
row_number() over(partition by dbid, instance_number, sample_time, event order by 1) r1
from t_ash t
/*where sample_time >
to_timestamp('2013-11-17 13:59:00',
'yyyy-mm-dd hh24:mi:ss')
and sample_time <
to_timestamp('2013-11-17 14:10:00',
'yyyy-mm-dd hh24:mi:ss')*/
) t
where r1 = 1) t
where r < 3
order by dbid, instance_number, sample_time, r;
SAMPLE_ID SAMPLE_TIME INSTANCE_NUMBER EVENT SESSION_STATE SESSION_COUNT
36402900 22:02:50.985 1 ON CPU 3
36402900 22:02:50.985 1 db file sequential read WAITING 1
36402910 22:03:01.095 1 ON CPU 1
36402920 22:03:11.195 1 db file parallel read WAITING 1
36402930 22:03:21.966 1 cursor: pin S wait on X WAITING 11
36402930 22:03:21.966 1 latch: shared pool WAITING 4
36402940 22:03:32.116 1 cursor: pin S wait on X WAITING 83
36402940 22:03:32.116 1 SGA: allocation forcing component growth WAITING 16
36402950 22:03:42.226 1 cursor: pin S wait on X WAITING 161
36402950 22:03:42.226 1 SGA: allocation forcing component growth WAITING 17
36402960 22:03:52.326 1 cursor: pin S wait on X WAITING 177
36402960 22:03:52.326 1 SGA: allocation forcing component growth WAITING 20
36402970 22:04:02.446 1 cursor: pin S wait on X WAITING 204
36402970 22:04:02.446 1 SGA: allocation forcing component growth WAITING 20
36402980 22:04:12.566 1 cursor: pin S wait on X WAITING 219
36402980 22:04:12.566 1 SGA: allocation forcing component growth WAITING 20
36402990 22:04:22.666 1 cursor: pin S wait on X WAITING 236
36402990 22:04:22.666 1 SGA: allocation forcing component growth WAITING 20
36403000 22:04:32.846 1 cursor: pin S wait on X WAITING 265
36403000 22:04:32.846 1 SGA: allocation forcing component growth WAITING 20
36403010 22:04:42.966 1 enq: US - contention WAITING 69
36403010 22:04:42.966 1 latch: row cache objects WAITING 56
36403020 22:04:53.076 1 db file scattered read WAITING 1
36403020 22:04:53.076 1 db file sequential read WAITING 1
从以上输出我们可以发现问题期间最严重的等待为cursor: pin S wait on X,高峰期等待该event的session数达到了265个,其次为SGA: allocation forcing component growth,高峰期session为20个。
***注意:
1) 再次确认以上输出有无断档,是否有某些时间没有采样。
2) 注意那些session_state为ON CPU的输出,比较ON CPU的进程个数与您的OS物理CPU的个数,如果接近或者超过物理CPU个数,那么您还需要检查OS当时的CPU资源状况,
比如OSWatcher/NMON等工具,高的CPU Run Queue可能引发该问题,当然也可能是问题的结果,需要结合OSWatcher和ASH的时间顺序来验证。
其原理为通过dba_hist_active_sess_history. blocking_session记录的holder来通过connect by级联查询,找出最终的holder. 在RAC环境中,每个节点的ASH采样的时间很多情况下并不是一致的,
因此您可以通过将本SQL的第二段注释的sample_time稍作修改让不同节点相差1秒的采样时间可以比较(注意最好也将partition by中的sample_time做相应修改)。
该输出中isleaf=1的都是最终holder,而iscycle=1的代表死锁了(也就是在同一个采样点中a等待b,b等待c,而c又等待a,这种情况如果持续发生,那么尤其值得关注)。采用如下查询能观察到阻塞链。
col event for a35
set lines 1000
set pages 9999
select /*+ parallel 8 */
level lv,
connect_by_isleaf isleaf,
connect_by_iscycle iscycle,
t.dbid,
t.sample_id,
t.sample_time,
t.instance_number,
t.session_id,
t.sql_id,
t.session_type,
t.event,
t.session_state,
t.blocking_inst_id,
t.blocking_session,
t.blocking_session_status
from t_ash t
/*where sample_time >
to_timestamp('2013-11-17 13:55:00',
'yyyy-mm-dd hh24:mi:ss')
and sample_time <
to_timestamp('2013-11-17 14:10:00',
'yyyy-mm-dd hh24:mi:ss')*/
start with blocking_session is not null
/*start with event = 'buffer busy waits'*/
connect by nocycle
prior dbid = dbid
and prior sample_time = sample_time
/*and ((prior sample_time) - sample_time between interval '-1'
second and interval '1' second)*/
and prior blocking_inst_id = instance_number
and prior blocking_session = session_id
and prior blocking_session_serial# = session_serial#
order siblings by dbid, sample_time;
LV ISLEAF ISCYCLE SAMPLE_TIME INSTANCE_NUMBER SESSION_ID SQL_ID EVENT SESSION_STATE BLOCKING_INST_ID BLOCKING_SESSION BLOCKING_SESSION_STATUS
1 0 0 22:04:32.846 1 1259 3ajt2htrmb83y cursor: WAITING 1 537 VALID
2 1 0 22:04:32.846 1 537 3ajt2htrmb83y SGA: WAITING UNKNOWN
***注意:
为了输出便于阅读,我们将等待event做了简写,下同。从上面的输出可见,在相同的采样点上(22:04:32.846),节点1 session 1259 在等待cursor: pin S wait on X,其被节点1 session 537 阻塞,
而节点1 session 537 又在等待SGA: allocation forcing component growth,并且ASH没有采集到其holder,因此这里cursor: pin S wait on X只是一个表面现象,
问题的原因在于SGA: allocation forcing component growth
比如如下SQL列出了每个采样点top 2的blocker session,并且计算了其最终阻塞的session数(参考blocking_session_count)
col event for a35
set lines 1000
set pages 9999
select t.lv,
t.iscycle,
t.dbid,
t.sample_id,
t.sample_time,
t.instance_number,
t.session_id,
t.sql_id,
t.session_type,
t.event,
t.seq#,
t.session_state,
t.blocking_inst_id,
t.blocking_session,
t.blocking_session_status,
t.c blocking_session_count
from (select t.*,
row_number() over(partition by dbid, instance_number, sample_time order by c desc) r
from (select t.*,
count(*) over(partition by dbid, instance_number, sample_time, session_id) c,
row_number() over(partition by dbid, instance_number, sample_time, session_id order by 1) r1
from (select
level lv,
connect_by_isleaf isleaf,
connect_by_iscycle iscycle,
t.*
from t_ash t
/*where sample_time >
to_timestamp('2013-11-17 13:55:00',
'yyyy-mm-dd hh24:mi:ss')
and sample_time <
to_timestamp('2013-11-17 14:10:00',
'yyyy-mm-dd hh24:mi:ss')*/
start with blocking_session is not null
/*start with event = 'buffer busy waits'*/
connect by nocycle
prior dbid = dbid
and prior sample_time = sample_time
/*and ((prior sample_time) - sample_time between interval '-1'
second and interval '1' second)*/
and prior blocking_inst_id = instance_number
and prior blocking_session = session_id
and prior
blocking_session_serial# = session_serial#) t
where t.isleaf = 1) t
where r1 = 1) t
where r < 3
order by dbid, sample_time, r;
SAMPLE_TIME INSTANCE_NUMBER SESSION_ID SQL_ID EVENT SEQ# SESSION_STATE BLOCKING_SESSION_STATUS BLOCKING_SESSION_COUNT
22:03:32.116 1 1136 1p4vyw2jan43d SGA: 1140 WAITING UNKNOWN 82
22:03:32.116 1 413 9g51p4bt1n7kz SGA: 7646 WAITING UNKNOWN 2
22:03:42.226 1 1136 1p4vyw2jan43d SGA: 1645 WAITING UNKNOWN 154
22:03:42.226 1 537 3ajt2htrmb83y SGA: 48412 WAITING UNKNOWN 4
22:03:52.326 1 1136 1p4vyw2jan43d SGA: 2150 WAITING UNKNOWN 165
22:03:52.326 1 537 3ajt2htrmb83y SGA: 48917 WAITING UNKNOWN 8
22:04:02.446 1 1136 1p4vyw2jan43d SGA: 2656 WAITING UNKNOWN 184
22:04:02.446 1 537 3ajt2htrmb83y SGA: 49423 WAITING UNKNOWN 10
22:04:12.566 1 1136 1p4vyw2jan43d SGA: 3162 WAITING UNKNOWN 187
22:04:12.566 1 2472 SGA: 1421 WAITING UNKNOWN 15
22:04:22.666 1 1136 1p4vyw2jan43d SGA: 3667 WAITING UNKNOWN 193
22:04:22.666 1 2472 SGA: 1926 WAITING UNKNOWN 25
22:04:32.846 1 1136 1p4vyw2jan43d SGA: 4176 WAITING UNKNOWN 196
22:04:32.846 1 2472 SGA: 2434 WAITING UNKNOWN 48
从以上输出,比如第一行,代表在22:03:32.116,节点1的session 1136 最终阻塞了82个session. 顺着时间往下看,可见节点1的session 1136 是问题期间最严重的holder,
它在每个采样点都阻塞了100多个session,并且它持续等待SGA: allocation forcing component growth,
注意观察其seq#您会发现该event的seq#在不断变化,表明该session并未完全hang住,由于时间正好发生在夜间22:00左右,这显然是由于自动收集统计信息job导致shared memory resize造成,
因此可以结合Scheduler/MMAN的trace以及dba_hist_memory_resize_ops的输出进一步确定问题。
**注意:
1) blocking_session_count 指某一个holder最终阻塞的session数,比如 a <- b<- c (a被b阻塞,b又被c阻塞),只计算c阻塞了1个session,因为中间的b可能在不同的阻塞链中发生重复。
2) 如果最终的holder没有被ash采样(一般因为该holder处于空闲),比如 a<- c 并且b<- c (a被c阻塞,并且b也被c阻塞),但是c没有采样,那么以上脚本无法将c统计到最终holder里,这可能会导致一些遗漏。
3) 注意比较blocking_session_count的数量与第3步查询的每个采样点的总session_count数,如果每个采样点的blocking_session_count数远小于总session_count数,那表明大部分session并未记载holder,因此本查询的结果并不能代表什么。
4) 在Oracle 10g中,ASH并没有blocking_inst_id列,在以上所有的脚本中,您只需要去掉该列即可,因此10g的ASH一般只能用于诊断单节点的问题。
除了通过ASH数据来找holder以外,我们还能用它来获取很多信息(基于数据库版本有所不同):
比如通过PGA_ALLOCATED列来计算每个采样点的最大PGA,合计PGA以分析ora-4030/Memory Swap相关问题;
通过TEMP_SPACE_ALLOCATED列来分析临时表空间使用情况;
通过IN_PARSE/IN_HARD_PARSE/IN_SQL_EXECUTION列来分析SQL处于parse还是执行阶段;
通过CURRENT_OBJ#/CURRENT_FILE#/CURRENT_BLOCK#来确定I/O相关等待发生的对象.
col event for a35
set lines 1000
set pages 9999
select /*+ parallel 8 */
level lv,
connect_by_isleaf isleaf,
connect_by_iscycle iscycle,
t.dbid,
t.sample_id,
t.sample_time,
t.instance_number,
t.session_id,
t.session_serial#,
t.session_type,
t.session_state,
t.SQL_OPNAME,
t.sql_id,
T.SQL_PLAN_HASH_VALUE,
t.SQL_PLAN_LINE_ID,
T.SQL_CHILD_NUMBER,
T.SQL_EXEC_ID,
T.SQL_EXEC_START,
T.IN_BIND,
t.IN_PARSE,
t.IN_HARD_PARSE,
t.IN_SQL_EXECUTION,
t.IN_PLSQL_COMPILATION,
t.IN_PLSQL_EXECUTION,
t.IN_SEQUENCE_LOAD,
T.IN_PLSQL_RPC,
t.PROGRAM,
t.MACHINE,
t.PGA_ALLOCATED,
t.TEMP_SPACE_ALLOCATED,
t.DELTA_READ_IO_BYTES,
t.DELTA_WRITE_IO_BYTES,
t.TM_DELTA_CPU_TIME,
t.event,
t.blocking_inst_id,
t.blocking_session,
t.blocking_session_serial#,
t.blocking_session_status,
t.WAIT_TIME,
t.TIME_WAITED,
t.CURRENT_OBJ#,
t.CURRENT_ROW#,
t.xid
from m_ash_202204162 t
start with event = 'buffer busy waits'
--start with blocking_session is not null
connect by nocycle prior dbid = dbid and prior blocking_session = session_id and prior blocking_session_serial# = session_serial#
and prior sample_time = sample_time and prior blocking_inst_id = instance_number order siblings by dbid, sample_time;
col event for a35
set lines 1000
set pages 9999
select /*+ parallel 4 */
t.dbid,
t.sample_id,
t.sample_time,
t.instance_number,
t.session_id,
t.session_serial#,
t.session_type,
t.session_state,
t.SQL_OPNAME,
t.sql_id,
T.SQL_PLAN_HASH_VALUE,
t.SQL_PLAN_LINE_ID,
T.SQL_CHILD_NUMBER,
T.SQL_EXEC_ID,
T.SQL_EXEC_START,
T.IN_BIND,
t.IN_PARSE,
t.IN_HARD_PARSE,
t.IN_SQL_EXECUTION,
t.IN_PLSQL_COMPILATION,
t.IN_PLSQL_EXECUTION,
t.IN_SEQUENCE_LOAD,
T.IN_PLSQL_RPC,
t.PROGRAM,
t.MACHINE,
t.PGA_ALLOCATED,
t.TEMP_SPACE_ALLOCATED,
t.DELTA_READ_IO_BYTES,
t.DELTA_WRITE_IO_BYTES,
t.TM_DELTA_CPU_TIME,
t.event,
t.blocking_inst_id,
t.blocking_session,
t.blocking_session_serial#,
t.blocking_session_status,
t.WAIT_TIME,
t.TIME_WAITED,
t.CURRENT_OBJ#,
t.CURRENT_ROW#,
t.xid
from m_ash_202204162 t
where session_id = 10501
order by dbid, instance_number,sample_time;
select * from dba_hist_sqltext where sql_id='bxgrjr8wa7g7x';
分析sql历史执行信息.md
原创声明:本文系作者授权腾讯云开发者社区发表,未经许可,不得转载。
如有侵权,请联系 cloudcommunity@tencent.com 删除。
原创声明:本文系作者授权腾讯云开发者社区发表,未经许可,不得转载。
如有侵权,请联系 cloudcommunity@tencent.com 删除。