Final post in the series, I hope to summarize my findings for everyone. Thanks for all the comments, feedback and comments I have received to-date on these posts.
Before providing the summary of all the testing, research and reading, I want to provide some internals for folks.
NUAM Memory Allocation
If a server has NUMA configuration, when we define the max memory setting. All memory gets allocated evenly from all NUMA nodes. For example, if I allocate 8GB from a server that has 4 NUMA nodes, it will allocate 2GB from each NUMA node for the instance in question. So be careful, if you are using CPU affinity. As even using CPU affinity, lets say you only want to use NUMA 1 for an instance; this can lead to foreign memory access. NUMA allocation still happens across all NUMA nodes (at least in SQL Server 2014). So even though you might only have subset of schedulers online, memory will still be accessed across all NUMA Nodes. Even though with new hardware foreign memory access does not have major impact on performance. Still recommend avoiding it. If you must use CPU affinity, then allocate sub-set of the processor from each NUMA node.
Lazy Writer Responsibility
Lazy Writer is back ground process and a thread is created per physical NUMA node (for Soft-NUMA, we only get one thread, you can read details here by Jonathan Kehayias (Blog|Twitter)). Lazy Writer responsibility is to make sure sufficient number of free pages, about 640 pages, are always available for new data to be loaded. If Lazy Writer thread is spinning, that is usually a clear signal of internal memory pressure.
Note: Internal Memory pressure does not mean server doesn’t have enough memory. There are multiple factors that can contribute to this, one fact is low Max Server Memory, which is what has been explored in these articles.
Original Scenario: SQL Server 2012 Latest Build + Windows 2008 R2 Latest Build + Locked Pages In Memory + Low Max Server Memory + AMD Processors + NUMA Configuration
In this build, I can consistently reproduce the issue. If I set the Max Memory low enough, it caused the Lazy Writer on NUMA node 0 to spin non-stop, as I documented in SQL Server 2012 NUMA Node Imbalance – Cont’d. To further to this I understand some additional internals around this thanks to colleague who helped here. As I was not true sure what is happening.
If we looked at the DBCC MEMORYSTATUS, in particular we notice following for Node 0:
Memory node Id = 0 KB
VM Reserved 127086340
VM Committed 642452
Locked Pages Allocated 55996
Pages Allocated 54008
Pages Free 0
Target Committed 512000
Current Committed 698448
Foreign Committed 0
Away Committed 0
Taken Away Committed 0
Review the two numbers high-lighted above, Target memory is what this NUMA node is allocated and Current is what is NUMA consuming, therefore the Page Free count is 0. Thus Lazy Writer spinning to free up memory. However it is unable, thus spinning indefinitely causing CPU to pin. Question to answer, why can’t Lazy Writer free up memory from NUMA Node 0?
Looking further into MEMORYSTATUS, we also notice, that most of the memory allocated into NUMA Node 0 is for buffer cache:
MEMORYCLERK_SQLBUFFERPOOL (node 0) KB
VM Reserved 103971820
VM Committed 524288
Locked Pages Allocated 2524
SM Reserved 0
SM Committed 0
Pages Allocated 8
Test Scenario #1: Windows 10 + SQL Server 2012 RTM + Low Max Server Memory + Intel Processors + No-NUMA (Physical)
In this, I was just trying to reproduce the issue on my desktop, with no luck. Memory allocation or such were nominal.
Test Scenario #2: Windows 2012 + SQL Server 2014 RTM + Low Max Server Memory + Intel Processors + No-NUMA (Hyper-V)
Again no issue, as per the title of the series, NUMA Imbalance. It was obvious and expected this two scenarios don’t show any issues. But I wanted to confirm it.
Test Scenario #3: SQL Server 2008 R2 + Windows 2008 R2 + Low Max Server Memory + AMD Processors
These test was on same servers as original scenario, however we did not notice any issue.
*** During testing, we realized an interesting fact about NUMA configuration in SQL Server 2008 R2. SQL Server swaps NUMA 0 and NUMA 1 CPU mapping when creating a logical mapping for schedulers. For example, if you have 2 NUMA Nodes each with 4 cores; SQL Server binds NUMA 0 – Processor 0 – 3 to Scheduler 4 – 7 and NUMA 1 – Processor 4 – 7 to Scheduler 0 – 3. This is considered NUMA Swap, documented in this MSDN blog article here by CSS Team. However this behavior does not seem to persist in SQL Server 2012.
Test Scenario #4: SQL Server 2014 + Windows 2012 R2 + Low Max Server Memory + Intel Processors + NUMA (Azure IaaS Virtual Machine)
This was interesting case, when I set the Max Server Memory to low value, the NUMA node 0’s processor time started to increase. However it was not as swear as SQL Server 2012, however these are different processors. I had same behavior as noticed above, Current Committed was higher then Target Committed, therefore NUMA thread on Node 0 was spinning. However interesting thing note, it was not as aggressive as SQL Server 2012. It was not continuously spinning, it allowed CPU to breath. However would wake up frequently as low as ever 20ms.
!!! Eliminated Locked Pages in Memory as a factor for NUMA Imbalance.
Test Scenario #5: SQL Server 2014 + Windows 2012 R2 + Low Max Server Memory + Intel Processors + Locked Pages In Memory + NUMA (Azure IaaS Virtual Machine)
Exact same behavior as above was noticed, however, after restart, SQL Server no longer had Current Committed higher then Target Committed. Even after lowering the max memory even further then Scenario #4. It did not cause as dramatic shift.
*** Another interesting thing to note in SQL Server 2014. Up-to SQL Server 2012, Lazy Writer Thread bound to CPU 0 of each NUMA Node. In SQL Server 2014, Lazy Writer Thread is binding to last CPU of each NUMA Node.
Conclusion & Things to Consider
- Configuring SQL Server Max Server Memory low memory will cause Lazy Writer to spin due to internal memory pressure issue.
- There is no value at which point this behavior starts, it depends on the load of the server. For example, in Original Scenario, due to server load under 4000MB it started spinning the Lazy Writer thread; however in Test Scenario #4, I had to go down to as much as 1000MB. Therefore on your servers look at Lazy Writer thread in conjunction with CPU and DBCC MEMORYSTATUS to understand what is truly happening.
- This behavior is consistent in whether I have Locked Pages in Memory or not.
- In Windows 2012 R2 NUMA allocation seem to be much better then Windows 2008 R2.
- SQL Server 2014 Lazy Writer thread does not seem to be as aggressive as SQL Server 2012.
- Few KB articles came up during my research in SQL Server 2012 that fix NUMA related issues, KB2819662 and KB2926223.
- There as been some fixes released for NUMA for Windows 2008 R2 operating system also, please review KB2155311 and KB2510206.
- In addition, if you are running a server with more then 64-logical cores, please review K-Group Configuration, for details please reference this article.
- Last but not least verify that memory is distributed evenly on physical server. That is the memory banks have been evenly allocated, i.e. you do not have more memory allocated in one bank versus another, as you might also see one NUMA node working extra harder if there is significant amount of foreign memory access.
Question to answer, why can’t Lazy Writer free up memory from NUMA Node 0?
SQL Server Lazy Writer is only responsible for clearing memory from Buffer Pool, it cannot remove or clear memory from any other memory clerk. Therefore, it is possible, that other internal components have memory allocated on NUMA Node 0 only, therefore NUMA Node 0 is unable to free up enough memory for buffer pool only. Looking at DBCC MEMORYSTATUS we can see which memory clerks allocate from NODE Node 0 only, there are few, for example, MEMORYCLERK_SQLCLR, MEMORYCLERK_SOSMEMMANAGER, OBJECTSTORE_LOCK_MANAGER, and MEMORYCLERK_XE_BUFFER all have allocation from NODE Node 0 only.