一、背景
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_ss_history的数据来分析问题的一种方法。
二、适用于
Oracle 10G或更高版本,本文适用于任何平台。
三、案例分析
在Oracle 10G中,我们引入了AWR和ASH采样机制,有一个视图gv$active_ssion_history会每秒钟将数据库所有节点的Active Session采样一次,而
dba_hist_active_ss_history则会将gv$active_ssion_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_ss_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。
1、Dump出问题期间的ASH数据:
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_ss_history创建一个新表m_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 ur/passwd
SQL> create table m_ash as lect * from dba_hist_active_ss_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 ur/passwd file=m_ash.dmp tables=(m_ash) log=p.log
然后导入到测试机:
$ imp ur/passwd file=m_ash.dmp log=m_ash.imp.log
2、验证导出的ASH时间范围:
t line 200 pages 1000
col sample_time for a25
col event for a40
alter ssion t nls_timestamp_format='yyyy-mm-dd hh24:mi:ss.ff';
lect /*+ parallel 8 */ //开启并行加快查询速度
t.dbid, t.instance_number, min(sample_time), max(sample_time), count(*) ssion_count
from m_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
分析: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。
3、确认问题发生的精确时间范围:
lect /*+ parallel 8 */
豆类食物有哪些
花样跳绳dbid, instance_number, sample_id, sample_time, count(*) ssion_count
from m_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
华北蝼蛄
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
2
2
2
分析:注意观察以上输出的每个采样点的active ssion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。
4、确定每个采样点的top n event:(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ssion_count指的是该采样点上的等待该event的ssion数量。)
lect t.dbid,t.sample_id,t.sample_time,t.instance_number,t.event,t.ssion_state,
t.c ssion_count
from (lect t.*,rank() over(partition by dbid, instance_number, sample_time order by c desc) r
from (lect /*+ parallel 8 */
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 m_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
人物传记
李成桂是中国人吗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 quential read WAITING 1
36402910 22:03:01.095 1 ON CPU 1
36402920 22:03:11.195 1 db file parallel read WAITING 1京剧脸谱歌词
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
******** 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 quential read WAITING 1
2
2
2
分析: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ssion数达到了265个,其次为
SGA: allocation forcing component growth,高峰期ssion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ssion_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。
5、观察每个采样点的等待链:
*原理:通过dba_hist_active_ss_history. blocking_ssion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b,b等c,而c又
等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:
断奶涨奶怎么办
lect /*+ parallel 8 */
level lv,
connect_by_isleaf isleaf,
connect_by_iscycle iscycle,
t.dbid,
t.sample_id,
t.sample_time,
t.instance_number,
t.ssion_id,
t.sql_id,
t.ssion_type,
t.event,
t.ssion_state,
t.blocking_inst_id,
t.blocking_ssion,
t.blocking_ssion_status
from m_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_ssion is not null
connect by nocycle
prior dbid = dbid
and prior sample_time = sample_time
/*and ((prior sample_time) - sample_time between interval '-1'
cond and interval '1' cond)*/
and prior blocking_inst_id = instance_number
and prior blocking_ssion = ssion_id
and prior blocking_ssion_rial# = ssion_rial#
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
2
2
分析:从上面的输出可见,在相同的采样点上(22:04:32.846),节点1 ssion 1259在等待cursor: pin S wait on X,其被节点1 ssion 537阻塞,而节点
1 ssion 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。
6、基于第5步的原理来找出每个采样点的最终top holder:
如下SQL列出了每个采样点top 2的blocker ssion,并且计算了其最终阻塞的ssion数(参考blocking
_ssion_count):
lect t.lv,
t.iscycle,
t.dbid,
t.sample_id,
t.sample_time,
t.instance_number,
t.ssion_id,
t.sql_id,
t.ssion_type,
t.event,
t.q#,
t.ssion_state,
t.blocking_inst_id,
t.blocking_ssion,
t.blocking_ssion_status,
t.c blocking_ssion_count
from (lect t.*,row_number() over(partition by dbid, instance_number, sample_time order by c desc) r
from (lect t.*, count(*) over(partition by dbid, instance_number, sample_time, ssion_id) c, row_number() over(partition by dbid,
instance_number, sample_time, ssion_id order by 1) r1
from (lect /*+ paralle
l 8 */
level lv,
connect_by_isleaf isleaf,
connect_by_iscycle iscycle,
t.*
from m_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_ssion is not null
connect by nocycle
prior dbid = dbid
and prior sample_time = sample_time
/*and ((prior sample_time) - sample_time between interval '-1'
cond and interval '1' cond)*/
and prior blocking_inst_id = instance_number
and prior blocking_ssion = ssion_id
and prior
blocking_ssion_rial# = ssion_rial#) 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
2
2
2
分析:注意以上输出,比如第一行,代表在22:03:32.116,节点1的ssion 1136最终阻塞了82个ssion。顺着时间往下看,可见节点1的ssion 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ssion,并且它持续等待SGA: allocation forcing component growth,注意观察其q#您会发现该
event的q#在不断变化,表明该ssion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。
春节安排注意:
1)blocking_ssion_count指某一个holder最终阻塞的ssion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ssion,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_ssion_count的数量与第3步查询的每个采样点的总ssion_count数,如果每个采样点的blocking_ssion_count数远小于总
ssion_count数,那表明大部分ssion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。