Will 64-bit increase the performance of my SQL Server application?
With 64-bit servers becoming more mainstream and SQL Server customers running more deployments on native 64-bit Windows there is a common misperception that running on 64-bit will always provide an increase in performance over 32-bit. Many customers are surprised to find that some workloads actually run faster on 32-bit SQL Server than on the native 64-bit version. The assumption that 64-bit will always outperform 32-bit is not true and is determined largely by characteristics of the particular workload and hardware. Workloads that have a large volume of small uniform transactions using relatively simple query plans (OLTP-like in nature) which are not experiencing memory pressure may not benefit from running on native 64-bit SQL Server.
Differences between 32-bit and 64-bit
There are differences between 32-bit and 64-bit processes in general which apply to SQL Server. The working set of 64-bit SQL Server is larger than 32-bit both in code and data; mainly due to doubling the size for pointers, 64-bit instructions, structures that are cacheline aligned and more. This means 64-bit will be more sensitive to the size of onboard chip caches which may be used less efficiently when compared to 32-bit, resulting in a higher Cost per Instruction (CPI). The end result is applications which are not memory constrained on 32-bit (especially OLTP) may incur a performance penalty when running on 64-bit using the same hardware. We have observed approximately 10-15% degradation in most cases however the impact will depend on the characteristics of the workload and hardware being used.
Which applications will benefit most from 64-bit?
Many SQL Server applications will realize significant performance benefit running on native 64-bit. This platform will also provide much better scalability in the long term so it is important not to be short sighted when evaluating minor performance differences between the two platforms. Native 64-bit SQL Server typically benefits memory hungry applications the most and also applications requiring memory for items other than data pages. Some of these include:
• Procedure Cache: Lots of stored procedures and dynamic SQL batches (OLTP)
• Workspace Memory: Large number of concurrent hash joins or group by’s; large-scale sorts and index builds (Data Warehouse)
• Connection memory: Large numbers of connections
• Thread memory: High Concurrency OLTP, Complex query plans on multi-CPU machines
• CLR GC Heap memory: CLR SP allocations
• Lock Memory: Large scale OLTP
SQL Server 32-bit version allows use of memory beyond the 2 or 3 GB (32-bit limitation of user mode virtual address space) through the use of Address Windowing Extensions however there are restrictions on the use of AWE mapped memory. Specifically, the items above can not utilize AWE mapped memory (on 32 bit). More on this below.
What to consider when deciding between 32-bit and 64-bit
When considering one platform over the other I would recommend thinking about the following:
1. Processor architecture plays a significant role. The processor used will impact the overhead introduced by running native 64-bit SQL Server on 64-bit Windows; specifically the amount of cache on the processor. Older generation processors with less L2/L3 cache will likely incur more of a penalty running native 64-bit SQL Server than will newer processors with larger amounts of cache. It is important to consider this and when possible favor chipsets that offer more cache. The performance gain offered by processors with larger cache sizes vary and can only be determined through testing. An exception to this rule would be AMD processors which have L2 cache but no L3 cache. The memory architecture of AMD processors is very different from Intel processors and the lack of L3 cache may be offset by the different architecture of the data transfer between processor and memory. It is important to note that we are not endorsing one processor over the other; we have observed excellent SQL Server performance on both Intel and AMD processors.
2. Memory latency will have a significant impact. The speed of data transfer between processors and memory will have a significant impact on performance. When accessing main memory latency is important for performance. Slower transfer speeds will result in higher latencies and a bigger performance penalty for cache misses. Faster transfer speed will have a positive, and often, significant impact on the performance of 64-bit SQL Server. Reducing memory latency is a key design point and strength of Non-Uniform Memory Access (NUMA) server architectures.
3. Consider the specifics of your particular workload. If your workload is running on 32-bit and is not experiencing memory pressure or does not have any of the requirements mentioned above then the workload will likely not benefit from 64-bit.
4. Look for unexpected behaviors when testing. We have seen a number of cases where moving to 64-bit resulted in unexpected behavior as a result of removing the virtual address space limitations. On 32-bit systems it is common to run with AWE enabled to allow SQL Server to utilize memory beyond 2 or 3 GB. This memory however can only be used for data cache. On 64-bit this restriction is lifted and non-data cache memory can grow much larger than was possible on 32-bit. This includes items such as procedure cache, lock memory, connection memory, workspace memory, etc… While beneficial in most cases there are also cases where this can expose issues with the application’s use of SQL Server. Some common scenarios include:
• Extensive growth of the procedure cache as a result of a high volume of non-parameterized SQL statements. Service Pack 2 for SQL 2005 has some optimizations to help address these issues but it is considered a best practice to parameterize queries when possible and the sp2 optimizations do not replace the need to follow this advice.
• Failure to put a limit on ‘max server memory’ resulting in system wide memory pressure. It is important on 64-bit systems to place an upper limit on ‘max server memory’ leaving sufficient room for the operating system and ensure the SQL Server service account is granted the ‘Lock pages in memory’ security privilege (On 64-bit SQL Server the lock pages in memory privilege is only valid for Enterprise edition). The goal of setting an upper limit is to ensure that there is sufficient memory for the OS, other applications on the machine and allocations within SQL Server that are satisfied from outside the buffer pool. SQL Server’s buffer pool is dynamic and will adjust to external memory pressure however being proactive is always a good strategy. It is not uncommon on systems with large amount of physical memory (>= 16GB) to leave several GB for these items. Refer to https://support.microsoft.com/kb/918483 for more information.
5. Think long term . Although there are times when a workload will perform slightly slower on native 64-bit the difference is most cases is not significant (usually 10% or less). It is important not to let a small impact to performance overshadow the long term benefits of 64-bit SQL Server. Native 64-bit will offer significantly better scalability than 32-bit and is the future of computing (especially for database applications). Even if a workload does not require or benefit from the additional virtual address space now it may in the future.
6. Monitor and Test. By monitoring current performance of your application you may be able to get an idea of whether or not your specific workload will benefit from 64-bit SQL Server. Some potential indicators of memory pressure which may indicate a scenario when 64-bit will help are:
• High rate of I/O coupled with low page life expectancy and/or constant lazy writer activity.
• Constant non-zero values for memory grants pending.
• High rate of temporary objects being created/destroyed.
• High value for stolen pages as a percentage of the total pages consumed by the SQL Server buffer pool.
Ultimately however the only way to definitely know whether or not your application will benefit from running native 64-bit is to test.