Manual:Performance tuning

Jump to: navigation, search
Language: English  • 한국어

This page provides a quick overview of ways to improve the performance of MediaWiki. See also Manual:CacheManual:Cache.

For the impatient[edit | edit source]

Note Note: Most of the tweaks have been collected in the puppet/modules/role/manifests/simple_performant.pp and puppet/modules/role/manifests/simple_miser.pp files; if you install puppet, you can make them all at once with a single command.

Cache[edit | edit source]

Opcode caching[edit | edit source]

See Manual:Cache#PHP caching and PHP configuration#Opcode caching

Opcode caches store the compiled output of PHP scripts, greatly reducing the amount of time needed to run a script multiple times. Supported opcode caches are APC, mmTurck, WinCache and XCache; see $wgMainCacheTypeManual:$wgMainCacheType.

memcached[edit | edit source]

See MemcachedMemcached

The user interface text and other expensive objects can be cached by the opcode user cache or memcached, as will logins and partially completed pages.

If you have enough available RAM, you should use memcached, which will require at least 80MB or more of RAM, about 60MB for code plus whatever you need for cache. If you balance your load across multiple webservers, you should use a dedicated memcached (cluster).

Output caching[edit | edit source]

See Manual:File cacheManual:File cache for instructions on enabling and configuring rendered page caching

MediaWiki pages can be computationally expensive to render. MediaWiki has an optional file caching system that stores the output of rendered pages. For larger sites, using an external cache like Squid or Varnish is preferable to using the file cache.

HTTP caching proxies and HTTP acceleration[edit | edit source]

See Manual:Squid cachingManual:Squid caching and Manual:Varnish cachingManual:Varnish caching

Simply put, HTTP accelerators/caching proxies (such as Squid and Varnish) store copies of pages sent out by the web server. When a cached page is requested, the accelerator serves up the copy instead of passing the response on to the web server. This can tremendously reduce the load on the web server. When a page is updated, the copy is removed from the accelerator's cache.

See also this article for instructions on using Apache's mod_cache_disk with MediaWiki.

Other web server tuning[edit | edit source]

Configuration settings[edit | edit source]

Large sites running MediaWiki 1.6 or later should set $wgJobRunRateManual:$wgJobRunRate to a low number, say 0.01. See Manual:Job queueManual:Job queue for more information.

PHP tuning[edit | edit source]

mbstring[edit | edit source]

Although MediaWiki can work without the mbstring PHP library, it is highly recommended for performance reasons (note: mbstring.func-overload configuration option must be off).

FastStringSearchFastStringSearch[edit | edit source]

Can have great impact in some cases, probably never harms. [1]

HHVM[edit | edit source]

HipHop Virtual Machine is a JIT for PHP developed by and used in production at Facebook. HHVM is not a magic bullet, but has favorable performance characteristics compared to Zend. HHVM support isn't complete in MediaWiki, and should not be attempted by the faint hearted (some brave attempts can be found at HipHop deploymentHipHop deployment).

Database configuration and setup[edit | edit source]

MySQL[edit | edit source]

For a heavy concurrent write load, InnoDB is essential. In MediaWiki 1.24 and earlier, set $wgAntiLockFlagsManual:$wgAntiLockFlags = ALF_NO_LINK_LOCK | ALF_NO_BLOCK_LOCK; to reduce lock contention, at the expense of introducing occasional inconsistencies. Use memcached, not the default MySQL-based object cache.

See below for some DB configuration tricks. You can also try and run the mysql-tuning-primer script to get some quick statistics and suggestions.

Multiple servers[edit | edit source]

The database software and web server software will start to fight over RAM on busy MediaWiki installations that are hosted on a single server. If your wiki has a consistent traffic, a logical step, once other performance optimizations have been made (and cache serves most of the content), is to put the database and web server on separate servers (or, in some cases, multiple separate servers, starting with a slave.) Also:

Benchmarking[edit | edit source]

Some tools can help quickly evaluate the effects of performance tuning.

See also[edit | edit source]

  1. 1.0 1.1 Niklas Laxström, Performance is a feature, December 9th, 2013.