How It Works: CMemThread and Debugging Them

The wait type of CMemThread shows up in outputs such as sys.dm_exec_requests. This post is intended to explain what a CMemThread is and what you might be able to do to avoid the waits. The easiest way to describe a CMemThread is to talk about a standard Heap, HeapCreate and the options (with or without HEAP_NO_SERIALIZE).
 
Serialization is the a process of making sure only one thread of execution can execute a specific segment of code. The technique is most often talked about when talking about Windows Synchronization objects, such as, Mutexes and CriticalSections.   
 
I think of it like the ticket dispenser. You get a ticket and wait your turn to be served. This just like a synchronization object, let’s look at an example.     
 

EnterCriticalSection // Wait for your turn

dwVal++ // Do something that no other thread is allowed to do unless they have the ticket

LeaveCriticalSeciton // Allow the next ticket owner to execute the code

 

While the example is simplistic it quickly applies to a Heap. To allocate memory from a heap you would use HeapAlloc. The heap maintains various lists that can only be adjusted by one thread at a time or it would corrupt the lists. Let’s take a closer look at a high level heap design.            

clip_image002
The heap can be made up of multiple segments (different ranges of memory) that are linked together and each segment can have used and free blocks of memory.
 
When a HeapAlloc takes place the heap will locate a free block to support the allocation, update the free list, update used information and could even allocate a new segment if necessary to create more free memory. The maintenance of the list(s) are important to making sure the heap structures properly remain intact. If multiple threads attempt to modify the heap structures, in parallel, the structures will become damaged and lead to memory corruption. (Scribblers: https://blogs.msdn.com/b/psssql/archive/2012/11/12/how-can-reference-counting-be-a-leading-memory-scribbler-cause.aspx)
 
When you create a heap with the HEAP_NO_SERIALIZE option your code must make sure you don’t make calls to HeapAlloc, HeapReAlloc, HeapFree, Heap* by more than one thread at a time. This is usually done using something like a CriticalSection around all Heap* invocations.
 

EnterCriticalSection HeapAlloc LeaveCriticalSection

EnterCriticalSection HeapFree         LeaveCriticalSection
 

If you allow the Heap to maintain synchronization it will provide an efficient synchronization wrapper on your behalf so you don’t have additional synchronization mechanisms in your code.

 

CMemObj

SQL Server has a class named CMemObj that can be thought of as acting like a heap for the SQL Server developers. Instead of using HeapCreate the developer is calls CreateMemoryObject (often called a PMO – pointer to memory object) that is backed by the SQL Server memory manager. If you execute a select against sys.dm_os_memory_objects you can see the various memory objects currently in use by the SQL Server. The CMemObj is responsible for handling common activities such as Alloc, Free, ReAlloc, … as you would expect.

 
Think of the CMemObj as a HEAP_NO_SERIALIZE option for the SQL Server developer. It is not thread safe so the memory object should only be used by a single thread.
 

CMemThread

The CMemThread is the serialization wrapper around a CMemObj. For example the CMemThread::Alloc looks like the following.

 

CMemThread::Alloc(…)

{

                Enter SOS_Mutex // CMEMTHREAD WAIT TYPE AND ACCUMULATION OCCURS HERE

                CMemObj::Alloc(…) // __super::Alloc

                Leave SOS_Mutex

}

 
The developer creates a memory object with the thread safe flag and SQL Server’s CreateMemoryObject will return a pointer to a CMemThread instead of the underlying CMemObj but overriding the necessary methods to provide the thread safe wrapper so the developer can share the memory object among any thread.     
 
When you get a CMEMTHREAD wait you are observing multi-threaded access to the same CMemObj causing a wait while another thread is completing Alloc, Free, …. This is to be expected as long as the wait does not become excessive. When the number of waits and wait time start to become significant it can indicate that you need to release the pressure on the specific memory object.
 

3 Types

There are 3 types of memory objects (Global, Per Numa Node, Per CPU). For scalability SQL Server will allow a memory object to be segmented so only threads on the same node or cpu have the same underlying CMemObj, reducing thread interactions from other nodes or cpus, thereby increasing performance and scalability.

 
Many of the SQL Server memory objects are already segmented by node or cpu and provide scalability. Reference the following post for more details: https://blogs.msdn.com/b/psssql/archive/2011/09/01/sql-server-2008-2008-r2-on-newer-machines-with-more-than-8-cpus-presented-per-numa-node-may-need-trace-flag-8048.aspx  
 

bThreadSafe = 0x2,

bPartitionedByCpu = 0x40,

bPartitionedByNode = 0x80, -T8048 upgrade from by Node to by CPU (Can’t upgrade from global to by CPU)

 
Looking at the creation_options in sys.dm_os_memory_objects you can determine if the memory object is partitioned and if so to what degree, node or cpu. If the object is not partitioned (global) the trace flag has no impact on upgrading the partitioning scheme.
 

Here is an example that shows the active memory objects that are partitioned by cpu.

 

select * from sys.dm_os_memory_objects

           where 0x40 = creation_options & 0x40

 

Will TF 8048 Help Reduce CMEMTHREAD Waits ?

Here is a query that you can run on your box when you see high CMEMTHREAD waits.

 

SELECT

type, pages_in_bytes,

CASE

WHEN (0x20 = creation_options & 0x20) THEN 'Global PMO. Cannot be partitioned by CPU/NUMA Node. TF 8048 not applicable.'

WHEN (0x40 = creation_options & 0x40) THEN 'Partitioned by CPU.TF 8048 not applicable.'

WHEN (0x80 = creation_options & 0x80) THEN 'Partitioned by Node. Use TF 8048 to further partition by CPU'

ELSE 'UNKNOWN'

END

from sys.dm_os_memory_objects

order by pages_in_bytes desc

 
If you see the top consumers being of type 'Partitioned by Node.’, you may use startup, trace flag 8048 to further partition by CPU. Note: Trace flag 8048 is a startup trace flag.
 

Removing Hot Memory Object

· If the memory object is NUMA partitioned you may be able to use the trace flag to further partition the object and increase performance.

· If the memory object is global or already partitioned by CPU you need to study and tune the queries impacting the memory object.

 
 

Troubleshooting

To troubleshoot this issue, we need to understand the code path that is causing contention on a memory object.

 
An example of this is the memory object used to track allocations for create table. The stack for which looks like the following:     
 

00 sqlservr!CMemThread<CMemObj>::Alloc

01 sqlservr!operator new

02 sqlservr!HoBtFactory::AllocateHoBt

03 sqlservr!HoBtFactory::GetFreeHoBt

04 sqlservr!HoBtFactory::GetHoBtAccess

05 sqlservr!HoBtAccess::Init

06 sqlservr!HoBtFactory::CreateHoBt

07 sqlservr!SECreateRowset

08 sqlservr!DDLAgent::SECreateRowsets

09 sqlservr!CIndexDDL::CreateRowsets

0a sqlservr!CIndexDDL::CreateEmptyHeap

 
Starting a workload of create table(s) can cause the specific memory object contention as shown in the following stack trace. 
 

00 ntdll!NtSignalAndWaitForSingleObject

01 KERNELBASE!SignalObjectAndWait

02 sqlservr!SOS_Scheduler::Switch

03 sqlservr!SOS_Scheduler::SuspendNonPreemptive

04 sqlservr!SOS_Scheduler::Suspend

05 sqlservr!EventInternal<Spinlock<154,1,0> >::Wait

06 sqlservr!SOS_UnfairMutexPair::LongWait

07 sqlservr!SOS_UnfairMutexPair::AcquirePair

08 sqlservr!CMemThread<CMemObj>::Alloc

09 sqlservr!operator new

0a sqlservr!HoBtFactory::AllocateHoBt

0b sqlservr!HoBtFactory::GetFreeHoBt

0c sqlservr!HoBtFactory::GetHoBtAccess

0d sqlservr!HoBtAccess::Init

0e sqlservr!HoBtFactory::CreateHoBt

0f sqlservr!SECreateRowset

10 sqlservr!DDLAgent::SECreateRowsets

11 sqlservr!CIndexDDL::CreateRowsets

12 sqlservr!CIndexDDL::CreateEmptyHeap

 
The call to sqlservr!SOS_UnfairMutexPair::LongWait, from a memory object, results in the CMEMTHREAD wait. You can use the following query to see wait information related to sessions and requests.
 

select r.session_id,r.wait_type,r.wait_time,r.wait_resource

from sys.dm_exec_requests r

join sys.dm_exec_sessions s

on s.session_id=r.session_id and s.is_user_process=1

 

session_id wait_type wait_time wait_resource

---------- --------------- ----------- ---------------

54 NULL 0

55 NULL 0

56 CMEMTHREAD 17062

57 CMEMTHREAD 17062

58 CMEMTHREAD 17063

59 CMEMTHREAD 17063

60 CMEMTHREAD 17062

 
Use Extended Events and collect call stacks for all waits on CMEMTHREAD using an asynchronous bucketizer (or histogram in SQL Server 2012.)         
 

--First get the map_key for CMEMTHREAD wait type from the name-value pairs for all wait types stored in sys.dm_xe_map_values

--NOTE :- These map values are different b/w SQL Server 2008 R2 and 2012

select m.* from sys.dm_xe_map_values m

      join sys.dm_xe_packages p on m.object_package_guid = p.guid

where p.name = 'sqlos' and m.name = 'wait_types'

      and m.map_value = 'CMEMTHREAD'

 

/*

name object_package_guid map_key map_value

------------------------------------------------------------ ------------------------------------ ----------- ---------------

wait_types BD97CC63-3F38-4922-AA93-607BD12E78B2 186 CMEMTHREAD

*/

--Create an Extended Events session to capture callstacks for CMEMTHREAD waits ( map_key=186 on SQL Server 2008 R2) --Create an Extended Events session to capture callstacks for CMEMTHREAD waits ( map_key=186 on SQL Server 2008 R2)

 

IF EXISTS(SELECT * FROM sys.server_event_sessions WHERE name='XeWaitsOnCMemThread')

      DROP EVENT SESSION [XeWaitsOnCMemThread] ON SERVER

CREATE EVENT SESSION [XeWaitsOnCMemThread] ON SERVER

ADD EVENT sqlos.wait_info(

    ACTION(package0.callstack,sqlserver.session_id,sqlserver.sql_text)

WHERE (

            [wait_type]=(186)) --map_key for CMEMTHREAD on SQL Server 2008 R2)

            AND [opcode] = (1)

            AND [duration]> 5000 -- waits exceed 5 seconds

            )

ADD TARGET package0.asynchronous_bucketizer

(SET filtering_event_name=N'sqlos.wait_info',

     source_type=1,

     source=N'package0.callstack')

WITH (MAX_MEMORY=4096 KB,EVENT_RETENTION_MODE=ALLOW_SINGLE_EVENT_LOSS,

MAX_DISPATCH_LATENCY=5 SECONDS,MAX_EVENT_SIZE=0 KB,MEMORY_PARTITION_MODE=NONE,TRACK_CAUSALITY=OFF,STARTUP_STATE=OFF)

GO

 

--Create second Xevent session to generate a mini dump of all threads for the first two wait events catpured for CMEMTHREAD

IF EXISTS(SELECT * FROM sys.server_event_sessions WHERE name='XeDumpOnCMemThread')

      DROP EVENT SESSION [XeDumpOnCMemThread] ON SERVER

CREATE EVENT SESSION [XeDumpOnCMemThread] ON SERVER

ADD EVENT sqlos.wait_info(

    ACTION(sqlserver.session_id,sqlserver.sql_text,sqlserver.create_dump_all_threads)

WHERE (

            [wait_type]=(186)) --map_key for CMEMTHREAD on SQL Server 2008 R2)

            AND [opcode] = (1)

            AND [duration]> 5000 -- waits exceed 5 seconds

            AND package0.counter <=2 --number of times to generate a dump

            )

add target package0.ring_buffer

WITH (MAX_MEMORY=4096 KB,EVENT_RETENTION_MODE=ALLOW_SINGLE_EVENT_LOSS,

MAX_DISPATCH_LATENCY=5 SECONDS,MAX_EVENT_SIZE=0 KB,MEMORY_PARTITION_MODE=NONE,TRACK_CAUSALITY=OFF,STARTUP_STATE=OFF)

GO

 
--Start the sessions

ALTER EVENT SESSION [XeWaitsOnCMemThread] ON SERVER STATE=START

GO

ALTER EVENT SESSION [XeDumpOnCMemThread] ON SERVER STATE=START

GO

 

When you collect data using the histogram target, you can acquire the un-symbolized call stack using the following query.

 

SELECT

    n.value('(@count)[1]', 'int') AS EventCount,

    n.value('(@trunc)[1]', 'int') AS EventsTrunc,

    n.value('(value)[1]', 'varchar(max)') AS CallStack

FROM

    (SELECT CAST(target_data as XML) target_data

     FROM sys.dm_xe_sessions AS s

     JOIN sys.dm_xe_session_targets t

     ON s.address = t.event_session_address

      WHERE s.name = 'XeWaitsOnCMemThread'

     AND t.target_name = 'asynchronous_bucketizer') as tab

CROSS APPLY target_data.nodes('BucketizerTarget/Slot') as q(n)

 

EventCount EventsTrunc CallStack

--------------------------------------------------------------------------------------

1 0 0x0000000001738BD8

0x0000000000E53F8B

0x0000000000E541C1

0x0000000000E529B6

0x0000000000FBF22A

0x0000000000F763CB

0x0000000000E578C4

0x0000000000E56DFA

0x0000000000F86416

 
Symbolize the stack addresses to function/method names using the ln command (Windows Debugger) and public symbols against the dump that was captured, as shown below. Note: The mini-dump capture is important because it contains the image versions, locations and sizes at the time the XEL capture took place.
 

0:049> .sympath SRV*c:\symcache_pub*https://msdl.microsoft.com/download/symbols

Symbol search path is: SRV*c:\symcache_pub*https://msdl.microsoft.com/download/symbols

Expanded Symbol search path is: srv*c:\symcache_pub*https://msdl.microsoft.com/download/symbols

0:049> .reload /f sqlservr.exe

0:049> ln 0x0000000001738BD8

(00000000`00e5462c) sqlservr!XeSosPkg::wait_info::Publish+0xe2 | (00000000`00e5471c) sqlservr!SETaskSuspendingNotification

0:049> ln 0x0000000001738BD8;ln 0x0000000000E53F8B;ln 0x0000000000E541C1;ln 0x0000000000E529B6;;ln 0x0000000000FBF22A;ln 0x0000000000F763CB;ln 0x0000000000E578C4;ln 0x0000000000E56DFA;ln 0x0000000000F86416;ln 0x0000000000F7D922;ln 0x0000000000F87943;ln 0x0000000000F0083B;ln 0x0000000000F05D00

(00000000`00e5462c) sqlservr!XeSosPkg::wait_info::Publish+0xe2

(00000000`00e53d58) sqlservr!SOS_Scheduler::UpdateWaitTimeStats+0x286

(00000000`00e54174) sqlservr!SOS_Task::PostWait+0x4d

(00000000`00e52890) sqlservr!EventInternal<Spinlock<154,1,0> >::Wait+0x1b2

(00000000`00f7628c) sqlservr!SOS_UnfairMutexPair::LongWait+0x104

(00000000`00e577f4) sqlservr!SOS_UnfairMutexPair::AcquirePair+0x46

(00000000`00e57858) sqlservr!CMemThread<CMemObj>::Alloc+0x6c

(00000000`00e56ddc) sqlservr!operator new+0x1e

(00000000`00f7d930) sqlservr!HoBtFactory::AllocateHoBt+0xba

(00000000`00ef4a38) sqlservr!HoBtFactory::GetFreeHoBt+0x12a

Once you have a symbolized stack you have a better understanding of the memory, contention point as well as the command(s) that are contributing to the contention.   Using the trace flag or changing the query can remove the contention and improve SQL Server performance.

 

Co-Author: Special Thanks and XEvent assistance Provided by: Rohit Nayak
Bob Dorr - Principal SQL Server Escalation Engineer

Comments

  • Anonymous
    January 14, 2013
    The comment has been removed

  • Anonymous
    March 26, 2013
    Excellent!  Used this information to assist in diagnosing and remedying a spinlock convoy on two different systems - an AMD server and an Intel server.  Both systems were hitting 100% CPU utilized with very few logical reads.  Trace flag 8048 resolved the issue. In Bod Borr's post about trace flag 8048, he said: "Remember this requires more memory overhead..." Where would this overhead show up - in the operation of the memory clerks?  Is there something from the DMVs or perfmon that I can capture before and after to indicate the cost of trace flag 8048 on a given system? Cross-posted the question at Bob's post as well. blogs.msdn.com/b/psssql/archive/2011/09/01/sql-server-2008-2008-r2-on-newer-machines-with-more-than-8-cpus-presented-per-numa-node-may-need-trace-flag-8048.aspx

  • Anonymous
    December 04, 2013
    Is this an issues with SQL Server 2005 Enterprise SP2? Thanks!

  • Anonymous
    June 12, 2014
    Hi Bob , Query you mentioned to see if TF8048 is required uses column pages_in_bytes but this col is not in Sys.dm_os_memory_objects till 2012 so you need to change query accordingly.