SQL Server 2012 : Latches and Spinlocks - Monitoring Latches and Spinlocks

In terms of monitoring latches, you’ve already been introduced to some of the DMVs. There’s more to monitoring latches than just DMVs, though. Performance Monitor also offers useful information about latches, as do extended events, which also provide information about spinlocks. Another option is to use memory dumps, but those are not covered here.

DMVs

The DMVs covered earlier are a useful point of reference. You should familiarize yourself with the contents of sys.dm_os_wait_stats, sys.dm_os_latch_stats, and sys.dm_os_spinlock_stats, and be comfortable with the output they provide. In addition, sys.dm_os_waiting_tasks will display a list of any tasks that are currently waiting on a resource, providing a useful session_id column that can be used to hook into other useful DMVs for information about sessions and the like.

sys.dm_os_wait_stats

This DMV has five columns.

• wait_type
• waiting_tasks_count
• wait_time_ms
• max_wait_time_ms
• signal_wait_time_ms

The first three were described earlier. max_wait_time_ms shows the largest wait time for a single wait since the DMV was cleared. signal_wait_time_ms is less relevant for latches, although it does get used if threads hit spinlock barriers.

sys.dm_os_latch_stats

This DMV has four columns.

• latch_class
• waiting_requests_count
• wait_time_ms
• max_wait_time_ms

These columns have all been described earlier.

sys.dm_os_spinlock_stats

This DMV has six columns.

• name
• collisions
• spins
• spins_per_collision
• sleep_time
• backoffs

A collision is recorded when a spinlock tries to acquire a resource but finds it unavailable. As a result, the spinlock starts spinning. This increases the spins but the collision has already been recorded. Usefully, this DMV also provides a spins_per_collision column, saving the user from doing the calculation.

I’m sure you can imagine that the number of spins is potentially quite large. Let’s just say that it’s a good thing that this column is a bigint type, which handles numbers up to 19 digits long. I don’t think the correct technical term is actually “gazillions,” but it feels right when you take a look at this DMV on busy systems that have been up for a while.

The sleep_time and backoffs columns simply report the amount of time that has been spent sleeping on spinlocks, and the number of backoffs.

Performance Monitor

Performance Monitor provides several useful counters to keep an eye on. Figure 1 shows a typical screenshot containing the list of counters in the SQLServer:Latches category for a machine.

FIGURE 1

Table 1 describes these counters.

TABLE 1: Useful Performance Monitor Counters

These performance counter values are also available using the DMV sys.dm_os_performance_counters (see Figure 2) (code file Ch7Monitoring.sql):

FIGURE 2

SELECT *
FROM sys.dm_os_performance_counters
WHERE object_name LIKE '%Latches%';

Note that although the object_name field appears to end in the string ’Latches’, this field is actually stored as nchar(256), rather than nvarchar(256), so there is a large amount of whitespace at the end, and that last % is needed.

Extended Events

If you open the New Session Wizard for Extended Events from SQL Server 2012 Management Studio, you will reach a screen from which you select the events you wish to capture. After reaching this, first scroll the Event library section to reveal the Channel drop-down box. Then, as shown in Figure 3, check Debug, which is unchecked by default.

FIGURE 3

Now you can search for spinlock and latch to find a list of extended events related to these areas, as described in Table 2.

TABLE 2: Spinlock and Latch Extended Events

You should now be able to create an XE session collecting these events. Bear in mind that you would typically expect to see many more latch_acquire_time events occurring than the other event types, and you might not want to bother collecting them.