Wikimedia budget/2007/Q1/hardware

Rationale

Servers

For the growth of the next 3 months, we need at least one new rack of Apache application servers. If the growth is strong (like in April 2006), another rack of servers will be required (which will be added as contingency budget).

The increase of the squid farm by another 20 servers helps to serve our readers without adding even bigger numbers of application servers. They are more expensive than application servers because they have much more memory and disks. The extension by 20 servers will likely be sufficient for the next 6 months.

The increasing size of the wikis and the ever increasing number of changes makes it necessary to split the database farm again. Currently, there are two clusters: en.wikipedia.org, and the rest. A split of the "rest" into two farms will required in the upcoming months.

Lucene is our internal full text index. While we have to extend our Lucene capacities, we probably don't need new servers for it. Currently we're serving most of the Lucene requests from old 32bit servers with only one CPU and about 1GB of memory. The new apaches we're planning to buy are about 6-8 times stronger, so that we can reuse some of the first generation application servers as Lucene servers.

The setup of a new mail service for the Foundation and its staff is planned for Q1. To support this, a new mail server and redundant data storage is needed to provide a high level of availability.

Network

On the network side, infrastructure is needed to connect the new servers to our network. The application servers are connected to cheaper access switches, the databases will be connected directly to the core router, so that we need additional access switches and an extension of the core switch.

In our first cage, we have a serial port server that allows us to remotely manage our servers and other equipment even when they're not working properly. The new cage does not yet have such a server, making remote maintenance difficult, if not impossible, for machines in that cage. A 48 port console server will be sufficient for the current servers and a substantial growth. All new servers will have IPMI and possibly a service processor, so a console server is only required for other equipment.

After the upgrade of Kennisnet's bandwidth in Amsterdam, we're limited by our 1GB uplink. We're also short on network ports, some of our equipment is connected using a very old, near-to-dead, 10 MBps network Hub. The exact configuration of the new network switch requires further discussion about the future of our CDN, but will be within the budget given below. The new switch will be sufficient to support future growth.

Not included

The following includes only hardware costs. Hosting costs and bandwidth costs are not included.

What's needed

Servers
Application servers, proposal is to buy them "per rack", about 38 per rack, including an access switch and maybe a management switch. At least one rack.	120'000 USD
Squids, ~20	100'000 USD
DBs, 4 additonal servers to build a third DB cluster, splitting the non-enwiki cluster	32'000 USD
Lucene, 0 (use older servers)	0 USD
Mail servers	10'000 USD
Network hardware
access switches for the app servers (1 per rack)	7'000 USD
line cards for the DBs and other important misc servers	11'000 USD
A serial console server for cage 2, e.g. Avocent ACS48	4'000 USD
a new core switch for knams	30'000 USD
Storage/backup - use storage1/2 and amane	0 USD
Total	314'000 USD

Contingency

Servers
Application servers, proposal is to buy them "per rack", about 38 per rack, including an access switch and maybe a management switch. At least one rack.	120'000 USD
Network hardware
access switches for the app servers (1 per rack)	3'500 USD
Total	123'500 USD

Purchases

• /purchase 1