Wikimedia servers

From Meta, a Wikimedia project coordination wiki
Jump to navigation Jump to search
Wikimedia Meta-Wiki
This box: view · talk · edit

Other languages: de, es, fr, it, ja, ko, zh, uk

Wikipedia and the other Wikimedia projects are run from several racks full of servers. See also Wikimedia’s technical blog and phabricator.

System architecture[edit]

Network topology[edit]

The Network topology is described in "Network design" at Wikitech.

Software[edit]

Simplified overview of the employed software as of October 2015. (A very complex LAMP "stack")
  • Our DNS servers run gdnsd. We use geographical DNS to distribute requests between our four data centers (3x US, 1x Europe) depending on the location of the client.
  • We use Linux Virtual Server (LVS) on commodity servers to load balance incoming requests. LVS is also used as an internal load balancer to distribute MediaWiki requests. For back end monitoring and failover, we have our own system called PyBal.
  • For regular MediaWiki web requests (articles/API) we use Varnish caching proxy servers in front of Apache HTTP Server.
  • All our servers run either Debian or Ubuntu Server.
  • For distributed object storage we use Swift.
  • Our main web application is MediaWiki, which is written in PHP (~70 %) and JavaScript (~30 %).[1]
  • Our structured data is stored in MariaDB since 2013.[2] We group wikis into clusters, and each cluster is served by several MariaDB servers, replicated in a single-master configuration.
  • We use Memcached for caching of database query and computation results.
  • For full-text search we use Elasticsearch (Extension:CirrusSearch).
  • https://noc.wikimedia.org/ – Wikimedia configuration files.
Diagram showing flow of data between Wikipedia’s servers. Twenty database servers talk to hundreds of Apache servers in the backend; the Apache servers talk to fifty squids in the frontend.
Overview of system architecture
Wikimedia server racks at CyrusOne

Hosting[edit]

See also: w:Wikimedia Foundation#Technology  and wikitech:Clusters

As of May 2018, we have the following colocation facilities (each name is derived from an acronym of the facility’s company and an acronym of a nearby airport):

eqiad
Application services (primary) at Equinix in Ashburn, Virginia (Washington, DC area).
codfw
Application services (secondary) at CyrusOne in Carrollton, Texas (Dallas-Fort Worth area).
esams
Caching at EvoSwitch in Amsterdam, the Netherlands.
ulsfo
Caching at United Layer in San Francisco.

eqsin

Caching at Equinix in Singapore.

History[edit]

The backend web and database servers are in Ashburn, with Carrollton to handle emergency fallback in the future. Carrollton was chosen for this as a result of the 2013 Datacenter RfC. At EvoSwitch, we have a Varnish cache cluster and several miscellaneous servers. The Kennisnet location is now used only for network access and routing.

Ashburn (eqiad) became the primary data center in January 2013, taking over from Tampa (pmtpa and sdtpa) which had been the main data centre since 2004. Around April 2014, sdtpa (Equinix – formerly Switch and Data – in Tampa, Florida, provided networking for pmtpa) was shut down, followed by pmtpa (Hostway – formerly PowerMedium – in Tampa, Florida) in October 2014.

In the past we've had other caching locations like Seoul (yaseo, Yahoo!) and Paris (lopar, Lost Oasis); the WMF 2010–2015 strategic plan reach target states: "additional caching centers in key locations to manage increased traffic from Latin America, Asia and the Middle East, as well as to ensure reasonable and consistent load times no matter where a reader is located."

EvoSwitch and Kennisnet are recognised as benefactors for their in-kind donations. See the current list of benefactors.

A list of servers and their functions used to be available at the server roles page; no such list is currently maintained publicly (perhaps the private racktables tool has one). It used to be possible to see a compact table of all servers grouped by type on icinga, but this is no longer publicly available. The puppet configuration provides a pretty good reference for software what each server runs however.

Status and monitoring[edit]

You can check one of the following sites if you want to know if the Wikimedia servers are overloaded, or if you just want to see how they are doing.

If you are seeing errors in real time, visit #wikimedia-techconnect on irc.freenode.net. Check the topic to see if someone is already looking into the problem you are having. If not, please report your problem to the channel. It would be helpful if you could report specific symptoms, including the exact text of any error messages, what you were doing right before the error, and what server(s) are generating the error, if you can tell.

Energy use[edit]

The Sustainability Initiative aims at reducing the environmental impact of the servers by calling for renewable energy to power them.

The Wikimedia Foundation's servers are spread out in four colocation data centers in Virginia, Texas and San Francisco in the United States, and Amsterdam in Europe. As of May 2016, the servers use 222 kW, summing up to about 2 GWh of electrical energy per year. For comparison: An average household in the United States uses 11 MWh/year, the average for Germany is 3 MWh/year.

Only the few servers in Amsterdam run on renewable energy, the other use different conventional energy mixes. Overall, just 9% of Wikimedia Foundation data centers' energy comes from renewable sources, with the rest split evenly between coal, gas and nuclear power (34%, 28%, and 28%, respectively) . The bulk of the Wikimedia movement's electricity demand is in Virginia and Texas, which unfortunately have both very fossil fuel heavy grids.

Server name Data center location Provider Date

opened

Average

energy consumption (kW)

Energy sources Carbon footprint (CO2/year) Renewable

option and cost

eqiad Ashburn, VA

20146-20149 USA

Equinix (Website) February 2011 May 2016: 130

May 2015: 152

32% coal

20% natural gas

25% nuclear

17% renewable

1,040,000 lb = 520 short tons = 470 metric tons

= 0.32 * 130 kW * 8765.76 hr/yr * 2.1 lb CO2/kWh for coal

+ 0.20 * 130 kW * 8765.76 hr/yr * 1.22lb CO2/kWh for nat gas

+ 0.25 * 130 kW * 8765.76 hr/yr * 0 lb CO2/kWh for nuclear

+ 0.17 * 130 kW * 8765.76 hr/yr * 0 lb CO2/kWh for renewable

In 2015, Equinix made "a long-term commitment to use 100 percent clean and renewable energy". In 2017, Equinix renewed this pledge.
codfw Carrollton, TX

75007 USA

CyrusOne (Website) May 2014 May 2016: 77

May 2015: 70

23% coal

56% natural gas

6% nuclear

1% hydro/biomass/solar/other

14% wind (Oncor/Ercot)

790,000 lb = 400 short tons = 360 metric tons

= 0.23 * 77 kW * 8765.76 hr/yr * 2.1 lb CO2/kWh for coal

+ 0.56 * 77 kW * 8765.76 hr/yr * 1.22lb CO2/kWh for nat gas

+ 0.06 * 77 kW * 8765.76 hr/yr * 0 lb CO2/kWh for nuclear

+ 0.15 * 77 kW * 8765.76 hr/yr * 0 lb CO2/kWh for renewables

?
esams Haarlem

2031 BE Netherlands

EvoSwitch (Website) December 2008 May 2016: < 10

May 2015: 10

"a combination of wind power, hydro and biomass" 0 n.a.
ulsfo San Francisco, CA

94124 USA

UnitedLayer (Website) June 2012 May 2016: < 5

May 2015: < 5

25% natural gas

23% nuclear

30% renewable

6% hydro

17% unspecified (PG&E)

13,000 lb = 6.7 short tons = 6.1 metric tons (+ unspecified)

= 0.00 * 5 kW * 8765.76 hr/yr * 2.1 lb CO2/kWh for coal

+ 0.25 * 5 kW * 8765.76 hr/yr * 1.22lb CO2/kWh for nat gas

+ 0.23 * 5 kW * 8765.76 hr/yr * 0 lb CO2/kWh for nuclear

+ 0.36 * 5 kW * 8765.76 hr/yr * 0 lb CO2/kWh for hydro/renewable

+ 0.17 * 5 kW * 8765.76 hr/yr * ? lb CO2/kWh for unspecified

?
eqsin Singapore Equinix (Website) ? ? ? ? ?

See also[edit]

More hardware info[edit]

Admin logs[edit]

Offsite traffic pages[edit]

Long-term planning[edit]

Historical information[edit]

Information about other web sites[edit]

  • a b See MediaWiki analysis, MediaWiki WMF-supported extensions analysis.
  • "Wikipedia Adopts MariaDB — Wikimedia blog" (text/html). blog.wikimedia.org. Wikimedia Foundation, Inc. 2013-04-22. Retrieved 2014-07-20.