一切福田，不離方寸，從心而覓，感無不通。

SQLSERVER排查CPU占用高的情况

今天中午，有朋友叫我帮他看一下数据库，操作系统是Windows2008R2 ,数据库是SQL2008R2 64位

64G内存，16核CPU

硬件配置还是比较高的，他说服务器运行的是金蝶K3软件，数据库实例里有多个数据库

 

现象

他说是这几天才出现的，而且在每天的某一个时间段才会出现CPU占用高的情况

内存占用不太高，只占用了30个G

CPU占用100%

排查方向

 

一般排查都是用下面的脚本，一般会用到三个视图sys.sysprocesses ,dm_exec_sessions ,dm_exec_requests

1
2
3
4
5
USE master
GO
--如果要指定数据库就把注释去掉
SELECT * FROM sys.[sysprocesses] WHERE [spid]>50 --AND DB_NAME([dbid])='gposdb'
SELECT COUNT(*) FROM [sys].[dm_exec_sessions] WHERE [session_id]>50

看一下当前的数据库用户连接有多少

然后使用下面语句看一下各项指标是否正常，是否有阻塞，这个语句选取了前10个最耗CPU时间的会话

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
SELECT TOP 10
[session_id],
[request_id],
[start_time] AS '开始时间',
[status] AS '状态',
[command] AS '命令',
dest.[text] AS 'sql语句',
DB_NAME([database_id]) AS '数据库名',
[blocking_session_id] AS '正在阻塞其他会话的会话ID',
[wait_type] AS '等待资源类型',
[wait_time] AS '等待时间',
[wait_resource] AS '等待的资源',
[reads] AS '物理读次数',
[writes] AS '写次数',
[logical_reads] AS '逻辑读次数',
[row_count] AS '返回结果行数'
FROM sys.[dm_exec_requests] AS der
CROSS APPLY
sys.[dm_exec_sql_text](der.[sql_handle]) AS dest
WHERE [session_id]>50 AND DB_NAME(der.[database_id])='gposdb'  
ORDER BY [cpu_time] DESC

如果想看具体的SQL语句可以执行下面的SQL语句，记得在SSMS里选择以文本格式显示结果

1
2
3
4
5
6
7
8
--在SSMS里选择以文本格式显示结果
SELECT TOP 10
dest.[text] AS 'sql语句'
FROM sys.[dm_exec_requests] AS der
CROSS APPLY
sys.[dm_exec_sql_text](der.[sql_handle]) AS dest
WHERE [session_id]>50  
ORDER BY [cpu_time] DESC

模拟了一些耗CPU时间的动作

 

还有查看CPU数和user scheduler数和最大工作线程数，检查worker是否用完也可以排查CPU占用情况

1
2
3
4
--查看CPU数和user scheduler数目
SELECT cpu_count,scheduler_count FROM sys.dm_os_sys_info
--查看最大工作线程数
SELECT max_workers_count FROM sys.dm_os_sys_info

查看机器上的所有schedulers包括user 和system
通过下面语句可以看到worker是否用完，当达到最大线程数的时候就要检查blocking了
对照下面这个表
各种CPU和SQLSERVER版本组合自动配置的最大工作线程数
CPU数                 32位计算机                        64位计算机
<=4                     256                                   512
8                        288                                   576
16                       352                                   704
32                       480                                   960

1
2
3
4
5
6
7
8
SELECT
scheduler_address,
scheduler_id,
cpu_id,
status,
current_tasks_count,
current_workers_count,active_workers_count
FROM sys.dm_os_schedulers

 

如果SQLSERVER存在要等待的资源，那么执行下面语句就会显示出会话中有多少个worker在等待

结合[sys].[dm_os_wait_stats]视图，如果当前SQLSERVER里面没有任何等待资源，那么下面的SQL语句不会显示任何结果

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
SELECT TOP 10
[session_id],
[request_id],
[start_time] AS '开始时间',
[status] AS '状态',
[command] AS '命令',
dest.[text] AS 'sql语句',
DB_NAME([database_id]) AS '数据库名',
[blocking_session_id] AS '正在阻塞其他会话的会话ID',
der.[wait_type] AS '等待资源类型',
[wait_time] AS '等待时间',
[wait_resource] AS '等待的资源',
[dows].[waiting_tasks_count] AS '当前正在进行等待的任务数',
[reads] AS '物理读次数',
[writes] AS '写次数',
[logical_reads] AS '逻辑读次数',
[row_count] AS '返回结果行数'
FROM sys.[dm_exec_requests] AS der
INNER JOIN [sys].[dm_os_wait_stats] AS dows
ON der.[wait_type]=[dows].[wait_type]
CROSS APPLY
sys.[dm_exec_sql_text](der.[sql_handle]) AS dest
WHERE [session_id]&gt;50  
ORDER BY [cpu_time] DESC

比如我当前执行了查询SalesOrderDetail_test表100次，由于表数据非常多，所以SSMS需要把SQLSERVER执行的结果慢慢的取走，

造成了ASYNC_NETWORK_IO等待

1
2
3
4
USE [AdventureWorks]
GO
SELECT * FROM dbo.[SalesOrderDetail_test]
GO 100

 

问题源头

经过排查和这几天的观察情况，确定是某些表缺失索引导致，现在在这些表上增加了索引，问题解决

1
2
3
4
5
6
select * from t_AccessControl        --权限控制表权限控制
select * from t_GroupAccess            --用户组权限表用户组权限
select * from t_GroupAccessType        --用户组权限类表用户组权限类
select * from t_ObjectAccess        --对象权限表对象权限
select * from t_ObjectAccessType    --对象权限类型表对象权限类型
select * from t_ObjectType            --对象类型表对象类型

查询CPU占用高的语句

1
2
3
4
5
6
7
8
9
10
11
SELECT TOP 10
    total_worker_time/execution_count AS avg_cpu_cost, plan_handle,
    execution_count,
    (SELECT SUBSTRING(text, statement_start_offset/2 + 1,
       (CASE WHEN statement_end_offset = -1
          THEN LEN(CONVERT(nvarchar(max), text)) * 2
          ELSE statement_end_offset
       END - statement_start_offset)/2)
    FROM sys.dm_exec_sql_text(sql_handle)) AS query_text
FROM sys.dm_exec_query_stats
ORDER BY [avg_cpu_cost] DESC

查询缺失索引

1
2
3
4
5
6
SELECT
    DatabaseName = DB_NAME(database_id)
    ,[Number Indexes Missing] = count(*)
FROM sys.dm_db_missing_index_details
GROUP BY DB_NAME(database_id)
ORDER BY 2 DESC;
1
2
3
4
5
6
7
8
9
10
11
12
13
SELECT  TOP 10
         [Total Cost]  = ROUND(avg_total_user_cost * avg_user_impact * (user_seeks + user_scans),0)
         , avg_user_impact
         , TableName = statement
         , [EqualityUsage] = equality_columns
         , [InequalityUsage] = inequality_columns
         , [Include Cloumns] = included_columns
FROM        sys.dm_db_missing_index_groups g
INNER JOIN    sys.dm_db_missing_index_group_stats s
        ON s.group_handle = g.index_group_handle
INNER JOIN    sys.dm_db_missing_index_details d
        ON d.index_handle = g.index_handle
ORDER BY [Total Cost] DESC;

定位问题后，新建非聚集索引

1
2
3
4
5
6
7
CREATE NONCLUSTERED INDEX IX_t_AccessControl_F4 ON dbo.t_AccessControl
(
     FObjectType
)include([FUserID], [FAccessType], [FAccessMask]) WITH( STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
GO
drop index IX_t_AccessControl_F4 on t_AccessControl

CPU占用恢复正常

 

总结

从多次历史经验来看，如果CPU负载持续很高，但内存和IO都还好的话，这种情况下，首先想到的一定是索引问题，十有八九错不了。

注意文章开头贴出的客户机器负载情况图

Copyright © 2024 Longsheng All rights reserved.