Improved NUMA support

In cases where the buffer pool memory allocation was bigger than size of the node, system would start swapping already allocated memory even if there is available memory on other node. This is would happen if the default NUMA memory allocation policy was selected. In that case system would favor one node more than other which caused the node to run out of memory. Changing the allocation policy to interleaving, memory will be allocated in round-robin fashion over the available node. This can be done by using the mysqld_safe numa_interleave option.

Another improvement implemented is preallocating the pages in the buffer pool on startup with innodb_buffer_pool_populate variable. This forces NUMA allocation decisions to be made immediately while the buffer cache is clean.

It is generally recommended to enable all of the options together to maximize the performance effects on the NUMA architecture.

Version Specific Information

  • 5.6.11-60.3

    Improved NUMA support implemented. This feature was ported from Twitter’s MySQL patches.

System Variables

variable innodb_buffer_pool_populate
Command Line:Yes
Config File:Yes
Location:mysqld
Scope:Global
Dynamic:No
Variable Type:Boolean
Default Value:OFF
Range:ON/OFF

When this variable is enabled, InnoDB preallocates pages in the buffer pool on startup to force NUMA allocation decisions to be made immediately while the buffer cache is clean.

Command-line Options for mysqld_safe

variable flush_caches
Command Line:Yes
Config File:Yes
Location:mysqld_safe
Dynamic:No
Variable Type:Boolean
Default Value:OFF
Range:ON/OFF

When enabled this will flush and purge buffers/caches before starting the server to help ensure NUMA allocation fairness across nodes. This option is useful for establishing a consistent and predictable behavior for normal usage and/or benchmarking.

variable numa_interleave
Command Line:Yes
Config File:Yes
Location:mysqld_safe
Dynamic:No
Variable Type:Boolean
Default Value:OFF
Range:ON/OFF

When this option is enabled, mysqld will run with its memory interleaved on all NUMA nodes by starting it with numactl --interleave=all. In case there is just 1 CPU/node, allocations will be “interleaved” between that node.

© Copyright Percona LLC and/or its affiliates 2009-2014.
Except where otherwise noted, this documentation is licensed under the following license:
CC Attribution-ShareAlike 2.0 Generic
Created using Sphinx 1.2.2.
This documentation is developed in Launchpad as part of the Percona Server source code.
If you spotted innacuracies, errors, don't understood it or you think something is missing or should be improved, please file a bug.