Skip to content

SQL Server Wait Statistics/ Wait and Queues

June 9, 2015

Waits always happen, that is the way SQL Server’s scheduling system works.

A thread is using the CPU (called RUNNING) until it needs to wait for a resource/queue. Then it is moved to an unordered list of threads called SUSPENDED. In the meantime, the next thread on the FIFO (first-in-first-out) queue of threads waiting for the CPU (called being RUNNABLE) is given the CPU and becomes RUNNING. If a thread on the SUSPENDED list is notified that it’s resource is available, it becomes RUNNABLE and is put on the bottom of the RUNNABLE queue. Threads continue this clockwise movement from RUNNING to SUSPENDED to RUNNABLE to RUNNING again until the task is completed. You can see processes in these states using the sys.dm_exec_requests DMV.

SQL Server keeps track of the time that elapses between leaving the RUNNING state and becoming RUNNING again (called the “wait time”) and the time spent on the RUNNABLE queue (called the “signal wait time” – i.e. how long does the thread need to wait for the CPU after being signaled that its resource is available). We need to work out the time spent waiting on the SUSPENDED list (called the “resource wait time”) by subtracting the signal wait time from the overall wait time.

To calculate cumulative wait statistics using the sys.dm_os_wait_stats DMV

WITH [Waits] AS
        [wait_time_ms] / 1000.0 AS [WaitS],
        ([wait_time_ms] - [signal_wait_time_ms]) / 1000.0 AS [ResourceS],
        [signal_wait_time_ms] / 1000.0 AS [SignalS],
        [waiting_tasks_count] AS [WaitCount],
        100.0 * [wait_time_ms] / SUM ([wait_time_ms]) OVER() AS [Percentage],
        ROW_NUMBER() OVER(ORDER BY [wait_time_ms] DESC) AS [RowNum]
    FROM sys.dm_os_wait_stats
    WHERE [wait_type] NOT IN (
        N'BROKER_TASK_STOP',                N'BROKER_TO_FLUSH',
        N'CHKPT',                           N'CLR_AUTO_EVENT',
        N'CLR_MANUAL_EVENT',                N'CLR_SEMAPHORE',
        N'EXECSYNC',                        N'FSAGENT',
        N'HADR_TIMER_TASK',                 N'HADR_WORK_QUEUE',
        N'KSOURCE_WAKEUP',                  N'LAZYWRITER_SLEEP',
        N'LOGMGR_QUEUE',                    N'ONDEMAND_TASK_QUEUE',
        N'SERVER_IDLE_CHECK',               N'SLEEP_BPOOL_FLUSH',
        N'SLEEP_DBSTARTUP',                 N'SLEEP_DCOMSTARTUP',
        N'SLEEP_SYSTEMTASK',                N'SLEEP_TASK',
        N'WAITFOR',                         N'WAITFOR_TASKSHUTDOWN',
        N'XE_DISPATCHER_WAIT',              N'XE_TIMER_EVENT')
    AND [waiting_tasks_count] > 0
    MAX ([W1].[wait_type]) AS [WaitType],
    CAST (MAX ([W1].[WaitS]) AS DECIMAL (16,2)) AS [Wait_S],
    CAST (MAX ([W1].[ResourceS]) AS DECIMAL (16,2)) AS [Resource_S],
    CAST (MAX ([W1].[SignalS]) AS DECIMAL (16,2)) AS [Signal_S],
    MAX ([W1].[WaitCount]) AS [WaitCount],
    CAST (MAX ([W1].[Percentage]) AS DECIMAL (5,2)) AS [Percentage],
    CAST ((MAX ([W1].[WaitS]) / MAX ([W1].[WaitCount])) AS DECIMAL (16,4)) AS [AvgWait_S],
    CAST ((MAX ([W1].[ResourceS]) / MAX ([W1].[WaitCount])) AS DECIMAL (16,4)) AS [AvgRes_S],
    CAST ((MAX ([W1].[SignalS]) / MAX ([W1].[WaitCount])) AS DECIMAL (16,4)) AS [AvgSig_S]
FROM [Waits] AS [W1]
INNER JOIN [Waits] AS [W2]
    ON [W2].[RowNum] <= [W1].[RowNum]
GROUP BY [W1].[RowNum]
HAVING SUM ([W2].[Percentage]) - MAX ([W1].[Percentage]) < 95; -- percentage threshold
Reset the aggregated statistics: DBCC SQLPERF (N’sys.dm_os_wait_stats’, CLEAR);
Show wait statistics: DBCC SQLPERF (N’waitstats’)
  1. Queues: the resources that the threads are waiting for / the waiting resources of threads
  2. Waits: the SQL Server tracks of the threads which are waiting for the resources/queues
  3. What is the most worrying wait type?
    1. 505:CXPacket wait: Indicates parallelism, not necessarily a problem by itself as The coordinator thread in a parallel query always accumulates these waits However, the problem is If the parallel threads are not given equal amounts of work to do, or one thread blocks, the waiting threads will also accumulate CXPACKET waits, which will make them aggregate a lot faster. One thread may have a lot more to do than the others, and so the whole query is blocked while the long-running thread completes. If this is combined with a high number of PAGEIOLATCH_XX waits, it could be large parallel table scans going on because of incorrect non-clustered indexes, or a bad query plan. If neither of these are the issue, you might want to try setting MAXDOP to 4, 2, or 1 for the offending queries
    2. 304:PageIOLatch_EX wait: a thread is waiting for a data page to be read into the buffer pool from disk. It may indicate a bottleneck at the IO subsystem level, but why is the IO subsystem having so many reads? It could be buffer pool/memory pressure (i.e. not enough memory for the workload), a sudden change in query plans causing a large parallel scan instead of a seek, plan cache boat, or a number of other things. Don’t assume the root cause is the IO subsystem as it is just a result/ an appearance
    3. 275: ASYNC_NETWORK_IO wait:
    4. LCK_M_X wait: a thread is waiting to be granted an exclusive lock on something

Leave a Comment

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s

%d bloggers like this: