博客园  :: 首页  :: 新随笔  :: 联系 :: 订阅 订阅  :: 管理

PAGELATCH_EX Contention on 2:1:103

Posted on 2016-03-10 22:33  nzperfect  阅读(316)  评论(0编辑  收藏  举报

This blog post is meant to help people troubleshoot page latch contention on 2:1:103. If that’s what brought you to this page, then hopefully you find it useful. But first…

A textbook entitled "Don't Panic"

Initial Investigation

Last week I was asked to help tackle a production outage. Queries were slow enough that the system was considered unavailable. And just like any other performance problem I started by having SQL Server tell me what was wrong with itself.

  • I first started with sp_whoisactive but it took about 30 seconds for it to return. Several queries were blocked by others and some of the lead blockers were waiting for a PAGELATCH_EX on 2:1:103
  • Then I checked on the most common waits using a query found at Paul Randal’s article: Wait statistics, or please tell me where it hurts. (I don’t have that page bookmarked, every time I’m in trouble, I google “Tell me where it hurts Paul Randal”). I learned that PAGELATCH_EX contention was our most serious wait type. Paul mentions that he sees this kind of contention most commonly on tempdb.
  • I followed a link from Paul’s article to Robert Davis’s Breaking Down TempDB Contention. A script there helped me to discover that while there was lots of tempdb contention, it was all on page 2:1:103, which is not PFS, GAM or SGAM.

Finding More Info

So I looked for more help.

  • Web Search: My work colleague found and then pointed me to Latch waits on 2:1:103? You are probably creating too many temp tables in SQL Server by Microsoft’s Matt Wrock who faced the same problem. If you found this page because you’re facing the same issue then stop reading this article and go read Matt’s first. He explains what’s going on better than I could. Think of this article as a kind of sequel to his article.
  • StackOverflow: (serverfault in this case) has TempDB contention on sysmultiobjrefs SQL 2005. The best answer there is Matt pointing back to his blog article.
  • Microsoft Support who helped us out, but maybe not as quick as we’d like. To be fair, the turnaround time that we would have been happy with was measured in nanoseconds.

With that information it became clear to us that we were creating temp tables too often. And we were creating them in a way that made it impossible for SQL Server to cache. Did you know that? That SQL Server can cache temp tables? When a query is done with a temp table, SQL Server can truncate that table and give it to the next query to avoid having to create it again. Cached temp tables reduce tempdb contention including contention on this page. But as it turns out, SQL Server cannot cache temp tables from ad hoc queries.

But who creates temp tables that often? We did, just by using a table valued parameter in a parameterized query. Since SQL Server began supporting table valued parameters (introduced in SQL Server 2008), we have been gradually moving towards this practice in lieu of sending xml to be shredded.

… And More Info Including Some Other Links

I can’t help including some extra articles I found on 2:1:103 contention.

Robert Davis blogged about 2:1:103 contention in Tempdb Contention That Can’t Be Soothed. His advice is to remove statements in code like “SELECT … INTO”.  However, I believe that such statements contribute to PAGELATCH contention only if the statement is not part of a stored procedure (i.e. can’t be cached). I also think there are more common causes of 2:1:103 PAGELATCH contention than SELECT … INTO statements.

But Robert did link to a demo by Paul Randal (an absolutely amazing 5 minute demo linked from a 2011 issue of his newsletter). Paul tells us

  • That “the SQL Team knows about this. It’s a known issue. Hopefully something will be done about it in one of the future releases.”
  • But unlike SGAM or GAM contention, there’s absolutely nothing you can do about this to spread the contention around.

From a DBA point of view there’s not much that can be done. So as DB Developers, we have to find a workaround. But before I get to that, I want to mention some things that were not so useful.

… And Some Bad Info

There was some red herrings out there…

  • Add more tempdb files (nope, wrong kind of tempdb contention).
  • Many resources suggested enabling trace flag T1118. It’s a trace flag that eases tempdb contention, but not this kind.
  • Get faster tempdb disks? No, PAGELATCH contention is for in-memory copies of pages, not disk (that’s PAGEIOLATCH).
  • Tune the queries in question? Not a bad idea, but this problem is about the number of problem queries, not the performance of each.
  • A fix from Microsoft Support: Fix: Poor performance in SQL Server 2008 R2 when table-valued functions use many table variables. Oooh… so close, but we weren’t using table-valued functions. And the workarounds they list (disabling AUTO_UPDATE_STATISTICS) did not help. But maybe it might help you?
  • The Object:Created event. Whether traced with Profiler or collected with an Extended Events session, this event can report on created temp tables. Maybe I can use this event to tell me which queries are creating the tempdb tables. Nope, not this time! This event  has two drawbacks which make it useless for troubleshooting 2:1:103 contention:
    • The Object:Created event reports the creation of temp tables even when they’re cached (which don’t need a latch on 2:1:103).
    • The Object:Created event doesn’t report the declaration of any table variables (which may need a latch on 2:1:103).

Strategy: So What Do We Do?

Knowing is half the battle right? But that means we still have a lot of work to do. I’m going to recap what we know so far. We know that we have trouble when there are queries that:

  • are executed frequently
  • create temp tables (either explicitly or by declaring table valued parameters).
  • are not cached (Microsoft explains when temp tables are not cached. In my case, it was because they were ad-hoc queries)
  • require a page latch on 2:1:103.

There is a performance counter that can track the all of the above (except maybe for that last bullet). It’s called Temp Tables Creation Rate and it’s found in the perf counter category “General Statistics”. Now this is a metric you can trust. We found that a high temp table creation rate was tightly correlated to the trouble we were seeing. So when troubleshooting, look at this performance counter (and leave the “Object:Create” event alone).

So now what do we do? First, we must find the ad hoc queries that create these temp tables. Then, we have to put them in stored procedures so that the temp tables can be cached. Alternatively, we could reduce the need for creating them. It’s a workaround, but it’s what we’ve got.

Finding Such Queries Is Difficult

But here’s the hard part. In a high volume system, it’s difficult to identify exactly which queries are causing the most trouble. Microsoft support can go through tons of collected trace and performance data to try to find such queries, but it’s a long process. On our side, we looked at sys.dm_os_waiting_tasks:

Select *
From sys.dm_os_waiting_tasks

And we saw all the contention on 2:1:103, but when we tried to look up the SQL text for it

Select wt.*, st.text
From sys.dm_os_waiting_tasks wt
left join sys.dm_exec_requests er
    on er.session_id = wt.session_id
outer apply fn_get_sql(er.sql_handle) st
where wt.resource_description is not null

The text was often unavailable. Basically, I’m guessing maybe the dmv’s I was using weren’t quick enough to tell me which queries were suffering from (or causing) contention on 2:1:103. So I decided to look through the cache for query candidates that might create temp tables. here’s what I came up with. It’s not a comprehensive list and there might be false positives, but it might be enough to go on. If you know your applications well, you can tailor the filters below to something more relevant for you.

select cp.plan_handle, sql_handle, text, refcounts, usecounts
from sys.dm_exec_cached_plans cp
join sys.dm_exec_query_stats qs
    on cp.plan_handle = qs.plan_handle
cross apply sys.dm_exec_sql_text(qs.sql_handle) st
where (
       (
        st.text like N'%READONLY%'  -- possible table valued parameter (prepared queries)
        or st.text like N'%TABLE %' -- possible table creation
        or st.text like '%#%'       -- possible use of temp table
        or st.text like '%INTO%'    -- possible table creation
       )
       and objtype in ('prepared', 'ad hoc') -- temp tables not cachable with these
)
order by usecounts desc

Going Forward

Personally after helping identify and implement the workarounds. I’m doing a couple things:

  • I’m recommending that developers not create temp tables or declare table variables that cannot be cached. For now, this means we use stored procedures for any query that uses table variables or temp tables.
  • We now have thresholds on our performance tests which look at the performance counter Temp Tables Creation Rate.
  • I created a Microsoft Connect item. If you’re troubleshooting the same problem, head over there and let Microsoft know you’re having trouble too.