A Context-Based Request Dispatching Policy for Layer-7 Web Cluster

Abstract

Locality-aware request distribution (LARD) [2] is a dispatching policy commonly used in the front-end node of the Web cluster. The principle of LARD is to dispatch all requests for identical context to the same back-end server. Such principle increases the likelihood to nd the requested context in the cache of the back-end server. However, LARD makes decision based on the number of TCP connections between the front-end node and each back-end server. Due to each connection does not always produce requests from clients, the decision made by LARD would not be totally accurate. In this paper, We propose an ecient context-based request dispatching policy to solve the problem of LARD. With our policy, back-end servers would send some state information such as input queue length and total service time of queue, back to the front-end node periodically. Obviously, the load of back-end server calculated based on feedbacks is more accurate than that only based on the numbers of TCP connections. In addition, we also develop two algorithms for packet classier and admission control mechanism, respeclity. These two mechanisms are used to mantain performance for Web cluster and avoid Web cluster overloaded. The simulation result demonstrates that our policy can achieve high cache hit rates and good load balance on the back-end servers, and uses smaller number of table entries in the front-end node than LARD.