SQL Server and Hyper-V Dynamic Memory Part 2
Part 1 of this series was an introduction and overview of Hyper-V Dynamic Memory. This part looks at SQL Server memory management and how the SQL engine responds to changing OS memory conditions.
Part 2: SQL Server Memory Management
As with any Windows process, sqlserver.exe has a virtual address space (VAS) of 4GB on 32-bit and 8TB in 64-bit editions. Pages in its VAS are mapped to pages in physical memory when the memory is committed and referenced for the first time. The collection of VAS pages that have been recently referenced is known as the Working Set.
How and when SQL Server allocates virtual memory and grows its working set depends on the memory model it uses. SQL Server supports three basic memory models:
1. Conventional Memory Model
The Conventional model is the default SQL Server memory model and has the following properties:
- Dynamic - can grow or shrink its working set in response to load and external (operating system) memory conditions.
- OS uses 4K pages – (not to be confused with SQL Server “pages” which are 8K regions of committed memory).
- Pageable - Can be paged out to disk by the operating system.
2. Locked Page Model
The locked page memory model is set when SQL Server is started with "Lock Pages in Memory" privilege*. It has the following characteristics:
- Dynamic -can grow or shrink its working set in the same way as the Conventional model.
- OS uses 4K pages
- Non-Pageable – When memory is committed it is locked in memory, meaning that it will remain backed by physical memory and will not be paged out by the operating system.
A common misconception is to interpret "locked" as non-dynamic. A SQL Server instance using the locked page memory model will grow and shrink (allocate memory and release memory) in response to changing workload and OS memory conditions in the same way as it does with the conventional model.
This is an important consideration when we look at Hyper-V Dynamic Memory – “locked” memory works perfectly well with “dynamic” memory.
* Note in “Denali” (Standard Edition and above), and in SQL 2008 R2 64-bit (Enterprise and above editions) the Lock Pages in Memory privilege is all that is required to set this model. In 2008 R2 64-Bit standard edition it also requires trace flag 845 to be set, in 2008 R2 32-bit editions it requires sp_configure 'awe enabled' 1.
3. Large Page Model
The Large page model is set using trace flag 834 and potentially offers a small performance boost for systems that are configured with large pages. It is characterized by:
- Static - memory is allocated at startup and does not change.
- OS uses large (>2MB) pages
The large page model is supported with Hyper-V Dynamic Memory (and Hyper-V also supports large pages), but you get no benefit from using Dynamic Memory with this model since SQL Server memory does not grow or shrink. The rest of this article will focus on the locked and conventional SQL Server memory models.
When does SQL Server grow?
For “dynamic” configurations (Conventional and Locked memory models), the sqlservr.exe process grows – allocates and commits memory from the OS – in response to a workload. As much memory is allocated as is required to optimally run the query and buffer data for future queries, subject to limitations imposed by:
- SQL Server max server memory setting. If this configuration option is set, the buffer pool is not allowed to grow to more than this value. In SQL Server 2008 this value represents single page allocations, and in “Denali” it represents any size page allocations and also managed CLR procedure allocations.
- Memory signals from OS. The operating system sets a signal on memory resource notification objects to indicate whether it has memory available or whether it is low on available memory. If there is only 32MB free for every 4GB of memory a low memory signal is set, which continues until 64MB/4GB is free. If there is 96MB/4GB free the operating system sets a high memory signal. SQL Server only allocates memory when the high memory signal is set.
To summarize, for SQL Server to grow you need three conditions: a workload, max server memory setting higher than the current allocation, high memory signals from the OS.
When does SQL Server shrink caches?
SQL Server as a rule does not like to return memory to the OS, but it will shrink its caches in response to memory pressure. Memory pressure can be divided into “internal” and “external”.
- External memory pressure occurs when the operating system is running low on memory and low memory signals are set. The SQL Server Resource Monitor checks for low memory signals approximately every 5 seconds and it will attempt to free memory until the signals stop.
To free memory SQL Server does the following:
· Frees unused memory.
· Notifies Memory Manager Clients to release memory
o Caches – Free unreferenced cache objects.
o Buffer pool - Based on oldest access times.
The freed memory is released back to the operating system. This process continues until the low memory resource notifications stop.
- Internal memory pressure occurs when the size of different caches and allocations increase but the SQL Server process needs to keep its total memory within a target value. For example if max server memory is set and certain caches are growing large, it will cause SQL to free memory for re-use internally, but not to release memory back to the OS. If you lower the value of max server memory you will generate internal memory pressure that will cause SQL to release memory back to the OS.
Memory pressure handling has not changed much since SQL 2005 and it was described in detail in a blog post by Slava Oks.
Note that SQL Server Express is an exception to the above behavior. Unlike other editions it does not assume it is the most important process running on the system but tries to be more “desktop” friendly. It will empty its working set after a period of inactivity.
How does SQL Server respond to changing OS memory?
In SQL Server 2005 support for Hot-Add memory was introduced. This feature, available in Enterprise and above editions, allows the server to make use of any extra physical memory that was added after SQL Server started. Being able to add physical memory when the system is running is limited to specialized hardware, but with the Hyper-V Dynamic Memory feature, when new memory is allocated to a guest virtual machine, it looks like hot-add physical memory to the guest. What this means is that thanks to the hot-add memory feature, SQL Server 2005 and higher can dynamically grow if more “physical” memory is granted to a guest VM by Hyper-V dynamic memory.
SQL Server checks OS memory every second and dynamically adjusts its “target” (based on available OS memory and max server memory) accordingly.
In “Denali” Standard Edition will also have sqlserver.exe support for hot-add memory when running virtualized (i.e. detecting and acting on Hyper-V Dynamic Memory allocations).
How does a SQL Server workload in a guest VM impact Hyper-V dynamic memory scheduling?
When a SQL workload causes the sqlserver.exe process to grow its working set, the Hyper-V memory scheduler will detect memory pressure in the guest VM and add memory to it. SQL Server will then detect the extra memory and grow according to workload demand. In our tests we have seen this feedback process cause a guest VM to grow quickly in response to SQL workload - we are still working on characterizing this ramp-up.
How does SQL Server respond when Hyper-V removes memory from a guest VM through ballooning?
If pressure from other VM's cause Hyper-V Dynamic Memory to take memory away from a VM through ballooning (allocating memory with a virtual device driver and returning it to the host OS), Windows Memory Manager will page out unlocked portions of memory and signal low resource notification events. When SQL Server detects these events it will shrink memory until the low memory notifications stop (see cache shrinking description above).
This raises another question. Can we make SQL Server release memory more readily and hence behave more "dynamically" without compromising performance? In certain circumstances where the application workload is predictable it may be possible to have a job which varies "max server memory" according to need, lowering it when the engine is inactive and raising it before a period of activity. This would have limited applicaability but it is something we're looking into.
What Memory Management changes are there in SQL Server “Denali”?
In SQL Server “Denali” (aka SQL11) the Memory Manager has been re-written to be more efficient. The main changes are summarized in this post. An important change with respect to Hyper-V Dynamic Memory support is that now the max server memory settingincludes any size page allocations and managed CLR procedure allocations it now represents a closer approximation to total sqlserver.exe memory usage. This makes it easier to calculate a value for max server memory, which becomes important when configuring virtual machines to work well with Hyper-V Dynamic Memory Startup and Maximum RAM settings.
Another important change is no more AWE or hot-add support for 32-bit edition. This means if you're running a 32-bit edition of Denali you're limited to a 4GB address space and will not be able to take advantage of dynamically added OS memory that wasn't present when SQL Server started (though Hyper-V Dynamic Memory is still a supported configuration).
In part 3 we’ll develop some best practices for configuring and using SQL Server with Dynamic Memory.