Abstract:

The replication of servers provides clients with the dilemma of choosing which server best suits its needs. Assuming all servers to be adequately replicated, this problem may be stated as how to choose the server capable of responding the fastest. The total response time will comprise server response time as well as network latency and load. When designing a novel application, logic may be provided in the clients allowing then to choose the best server, however existing application must function with its current installed client base.

The server selection method must only take a fraction of the total session time in order to be of any benefit. This seriously restricts the complexity of algorithms used in short lived application such as HTTP.

This paper presents a server selection technique performed by the network layer, resorting to an anycast address and a, optionally QoS enabled, routing protocol. This technique, named anycast server selection, combines network proximity (or other QoS metric) with CPU utilisation in order to select the best server.

The performance of this technique is compared, by means of simulations, to that of widely used techniques such as random and round robin server selection, and ideal techniques such as choosing the instantly least loaded server.

Date: 2005-Oct-20     Time: 14:00:00     Room: IST – TAGUS PARK SALA – 2.N.7.1

For more information:

• scosta@posi.ist.utl.pt
• 213100338