Web Farms

When a service grows beyond the capabilities of a single machine, groups of machines are often employed to provide the service. In the case of HTTP² and HTTPS³servers, or web servers, this is often referred to a Web Farm. Web farms typically employ both high availability and scalability technologies in order to provide a highly available service spread across multiple machines, with a single point of contact for clients.

Figure 1: Sample Web Farm

Web farms can take many forms, however, the three tiered approach is a useful model for explaining how a web farm works. A sample web farm is shown in figure 1. [2]

• The top layer of servers handles the multiplexing of incoming clients to web servers. In the case of the example given, incoming traffic travels through the router to the active IPVS Server -- the other IPVS server is a hot stand-by. The IPVS server then routes the client to one of the back end web servers.

  A similar topology would be implemented for other host-based layer 4 switching technologies such as the Cisco LocalDirector or the F5 BIG/ip. It is possible to construct a topology whereby the layer 4 switching servers are also the gateway routers to the network. However, it is often desirable to separate routing and multiplexing functionality to provide greater flexibility in how traffic is handled. If a layer 4 switch is used then the IPVS servers are eliminated and this switch forms all or part of the switching fabric for the Server Network.

• The middle layer contains the web servers. Typically, this layer contains the largest number of servers and these servers contain no content and very little state. These servers can be thought of as shared disk or network file system or RDBMS⁴ to HTTP or HTTPS converters. If any complex processing of requests is to be done then it should be done here as the compute power of this layer can easily be increased by adding more servers. Where possible state should be stored on clients by either using cookies or encoding the state into the URL⁵. This prevents the necessity for a client to repeatedly connect to the same server which makes for more flexible load-balancing, enabling a client session to continue even if a server fails. This is the layer where Linux servers are most likely to be found today.

• The bottom layer contains the data or truth source. There are many options here and common choices include servers for network file systems such as NFS⁶ and AFS⁷ or Database Servers for RDBMSs such as Oracle, MySQL, mSQL and PostgreSQL. In the future server independent storage such as that supported by GFS⁸ are likely to be utilised in this layer.

If a geographically distributed system is required then intelligent DNS solutions such as Resonate and Eddieware can be employed that will return the IP address of one of the servers, based on some heuristic. Alternatively, a central web server can handle all incoming requests and distribute them using an HTTP redirect after making a decision on which server the client should be directed to. The rewrite module that ships with the Apache HTTP Server is a very useful method of achieving this. It is also possible, using EBGP4⁹, to advertise the same network in more than once place and let the routing topology route customers to the most appropriate web server for them. Of course any instance of a web server in this discussion can be replaced with a web farm as per figure 1. A web farm of web farms if you will.