4.4. Dynamic Load Balancing

With many sites getting millions of hits per day, traffic to a Web server may get too high for one computer to handle it effectively. The obvious solution to servicing a heavy traffic load is to put more servers to work. There are a number of techniques to split the traffic across servers, but three common methods are DNS-based (Domain Name Server), dispatcher-based, and server-based [4].

The latter is a technique by which the server responds to a client with a new server address to which the client will resubmit the request. The redirected URL could be on the same computer as the main server or any one of several back-end mirror computers. The main server redirects the traffic to back-end Web servers according to some load balancing mechanism. Although the technique is transparent to users, it adds an extra connection to the original request and may, in some cases, increase the user response time and network traffic on the Internet or intranets. The DNS service translates a domain name to an IP address. Thus, when the DNS service receives a mapping request, it selects the IP address of one of the servers in the cluster. In round-robin DNS systems, the Web server name is associated with a list of IP addresses. Each IP address on the list maps to a different server, and each server contains a mirrored version of the website or access to a common file system. When a request is received, the Web server name is translated to the next IP address on the list. By translating Web server names to IP addresses in a round-robin fashion, this technique tries to balance the load among the servers. Another technique, called single-IP-image, has been used for distributing HTTP requests for a single service to different machines in a cluster. Basically, the technique uses the address of a special TCP router as the single address of the Web cluster. Client requests are addressed to the router that, by its turn, dispatches them to a server, according to some scheduling rules (i.e., server load characteristics). To make the dispatching transparent to users, the selected server returns the response with the router address, rather than its own address.