缓解 SQL Server has encountered 727 occurrence(s) of I/O requests taking longer than 15 seconds

sql server 会记录IO等待时间超过15 seconds的请求，这时application会有 time out 现象，dba须要判断是workload,concurrecy 所致仍是sql server配置没有最优致使？两者之间的比例各占多少？

1：判断sql server的配置是否最优化，相对容易，能够有如下项目检查：

　　a: 经过datafile/logfile放置在不一样的drive上，能够分离random io 和 sequence io 操做，能够极大的缓解读写IO，对于写IO，只需当即作logfile 的 sequence io便可，至于什么时候把data page 真正写入到datafile,

　　　　sql server本身把握。根据我本身的观察，分离datafile/logfile，一个月有个2到3次的 IO stall, 若是都在同一个磁盘，有个8到9次 IO stall 也很常见。

　　b: 尽可能合理的设置[max server memory (MB)], 通常为总内存的7/8便可。

　　c: tempdb最好在单独的磁盘上，若是不行的话，也要作datafile/logfile分离。

 

2：若是出现了问题，那如何判断是workload致使的呢？第一要作好IO activity的统计工做，第二要通知开发，是不是业务的变化致使IO增多，困难是这个IO stall 到底涉及到哪些表？如何获得？

　　a: 经过如下的sp, 利用sys.dm_io_virtual_file_stats DMV 记录每一分钟sql server IO读写操做

　　　　

USE [master]
GO
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
create proc [dbo].[usp_monitor_virtual_file_latency]
as
begin
 if OBJECT_ID('master.dbo.tbl_virtual_file_latency','U') is null
 begin
  create table master.dbo.tbl_virtual_file_latency
  (
   id int identity(1,1) primary key,
   collect_dt datetime not null,
   [database_name] [nvarchar](128) NULL,
   [file_id] [smallint] NOT NULL,
   [type_desc] [nvarchar](60) NULL,
   [sample_ms] [int] NOT NULL,
   [num_of_reads] [bigint] NOT NULL,
   [num_of_bytes_read] [bigint] NOT NULL,
   [io_stall_read_ms] [bigint] NOT NULL,
   [num_of_writes] [bigint] NOT NULL,
   [num_of_bytes_written] [bigint] NOT NULL,
   [io_stall_write_ms] [bigint] NOT NULL,
   [io_stall] [bigint] NOT NULL,
   [size_on_disk_bytes] [bigint] NOT NULL,
   [file_handle] [varbinary](8) NOT NULL,
   [physical_name] [nvarchar](260) NOT NULL,
   [state_desc] [nvarchar](60) NULL
  )
  create index idx_collect_dt on tbl_virtual_file_latency(collect_dt) with(online=on)
 end
 
 insert into master.dbo.tbl_virtual_file_latency
 SELECT 
    --virtual file latency
  GETDATE(),
  db_name(vfs.[database_id]) as database_name
      ,vfs.[file_id]
      ,mf.[type_desc]
      ,vfs.[sample_ms]
      ,vfs.[num_of_reads]
      ,vfs.[num_of_bytes_read]
      ,vfs.[io_stall_read_ms]
      ,vfs.[num_of_writes]
      ,vfs.[num_of_bytes_written]
      ,vfs.[io_stall_write_ms]
      ,vfs.[io_stall]
      ,vfs.[size_on_disk_bytes]
      ,vfs.[file_handle]
      ,mf.[physical_name]
      ,mf.[state_desc]

FROM sys.dm_io_virtual_file_stats (NULL,NULL) AS [vfs]
JOIN sys.master_files(nolock) AS [mf]
    ON [vfs].[database_id] = [mf].[database_id]
    AND [vfs].[file_id] = [mf].[file_id]
 WHERE DB_NAME(vfs.database_id) not in('master','model','msdb')
 
 
 --delete old data
 
 delete from master.dbo.tbl_virtual_file_latency where collect_dt <DATEADD(MONTH,-2,getdate())
end

 

 

 

　　

　　b: 经过如下的sp,求每分钟内每一个数据文件的IO的size和读写状况

　　　

USE [master]
GO
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
create proc [dbo].[usp_calculate_virtual_file_stats_PerMinutes]
@physical_name nvarchar(1000),
@collect_starttime datetime,
@collect_endtime datetime
as
with snap as
 (
 select
  row_number() over (order by id asc) as row_num,
  id
      ,[collect_dt]
      ,[database_name]
      ,[file_id]
      ,[type_desc]
      ,[sample_ms]
      ,[num_of_reads]
      ,[num_of_bytes_read]
      ,[io_stall_read_ms]
      ,[num_of_writes]
      ,[num_of_bytes_written]
      ,[io_stall_write_ms]
      ,[io_stall]
from master.dbo.tbl_virtual_file_latency(nolock) AS [vfs1]
WHERE vfs1.physical_name=@physical_name
and vfs1.collect_dt>=@collect_starttime
and vfs1.collect_dt<=@collect_endtime

)
select cur.collect_dt,
 read_io_num= case when (cur.num_of_reads-pre.num_of_reads)=0 then 0 else cur.num_of_reads-pre.num_of_reads end ,
 write_io_num= case when (cur.num_of_writes-pre.num_of_writes)=0 then 0 else cur.num_of_writes-pre.num_of_writes end ,
 read_latency_MS =case when (cur.num_of_reads-pre.num_of_reads)=0 then 0 else cast((cur.io_stall_read_ms-pre.io_stall_read_ms)*1./(cur.num_of_reads-pre.num_of_reads) as decimal(10,1)) end ,
 write_latency_MS= case when (cur.num_of_writes-pre.num_of_writes)=0 then 0 else cast((cur.io_stall_write_ms-pre.io_stall_write_ms)*1./(cur.num_of_writes-pre.num_of_writes) as decimal(10,1)) end ,
 average_latency_MS=case when (cur.num_of_reads-pre.num_of_reads+cur.num_of_writes-pre.num_of_writes)=0 then 0 else cast((cur.io_stall-pre.io_stall)*1./(cur.num_of_reads-pre.num_of_reads+cur.num_of_writes-pre.num_of_writes) as decimal(10,1)) end,
 averageKB_read=case when (cur.num_of_reads-pre.num_of_reads)=0 then 0 else cast((cur.num_of_bytes_read-pre.num_of_bytes_read)*1./(cur.num_of_reads-pre.num_of_reads)/1024 as decimal(10,1))end ,
 averageKB_write=case when (cur.num_of_writes-pre.num_of_writes)=0 then 0 else cast((cur.num_of_bytes_written-pre.num_of_bytes_written)*1./(cur.num_of_writes-pre.num_of_writes)/1024 as decimal(10,1)) end,
 averageKB_Transfer=case when (cur.num_of_reads-pre.num_of_reads+cur.num_of_writes-pre.num_of_writes)=0 then 0 else cast((cur.num_of_bytes_read-pre.num_of_bytes_read+cur.num_of_bytes_written-pre.num_of_bytes_written)*1./(cur.num_of_reads-pre.num_of_reads+cur.num_of_writes-pre.num_of_writes)/1024 as decimal(10,1)) end
from snap as pre join snap as cur
on pre.row_num=cur.row_num-1

 

 

　　

　　c: 最后根据历史数据来看IO读写的趋势

　　　　

exec usp_calculate_virtual_file_stats_PerMinutes 
'E:\TestDB\test.mdf',
'2014-12-14 13:49:01.000',
'2014-12-14 15:49:01.000'

 

　　　　

 

3：根据IO stall的信息找出sql server 读取哪一个表致使了IO Stall.

　　　　sql server log text: 

　　　　SQL Server has encountered 727 occurrence(s) of I/O requests taking longer than 15 seconds to complete  on file [F:\MSSQL\Test_1.ndf:MSSQL_DBCC7] in database [Test] (7). The OS file handle is 　　　　　　　　　　　　0x0000000000001A40. The offset of the latest long I/O is: 0x000023101f0000

 

　　　　

 　　　 a: 把偏移量0x000023101f0000转化为十进制为：150594322432

　　　　b: 把该值除以8192即为page number: 150594322432/8192=18383096

　　　　c:在经过page number求的table name, 再与开发讨论IO的问题.

　　　　

select object_name(object_id) as tblname,* 
from sys.partitions (nolock)as sp
where sp.partition_id in (
select au.container_id from 
sys.dm_os_buffer_descriptors(nolock) as obd
join sys.allocation_units(nolock) as au
on obd.allocation_unit_id=au.allocation_unit_id
where obd.page_id=18383096)

 

 　　　　

　　d: 若是该page number再也不内存中的化，该SQL 是没有结果的。可使用dbcc page(dbid,fileid,pageid,3)来获得该表的object_id.

 

结论：IO stall 只可能缓解，是无法避免的，业务量致使的IO永远在某一点上是超过存储的容量的，dba能经过各类事前过后方法到找到root cause, 据此提出解决的方法。但愿以上的方法能对你们有所启发。