• Summary
• Related Documents
  • Active Bibliography
  • Co-citation
• Version History

Abstract:

With the explosion in WWW traffic, Web server performance is becoming increasingly important. We consider a cluster-based network server in which all incoming requests are distributed among a number of backend machines. Specifically we make use of the content aware distribution of requests (as proposed in [1]). This scheme takes into account the content of the requested document to achieve high locality in the back end's main memory cache. We propose a new distributed architecture for a cluster based web server, with a switch based frontend. The switch performs a layer-four switching (Varghese et al [2]) and directs the incoming packets to one of the backends without regard to content. To make use of the locality in the main memory cache at the backends we introduce a centralized dispatcher. The backend which receives a request communicates with the centralised dispatcher to decide whether it or some other backend in the cluster should handle the request. Subsequently, it might handoff the TCP connection to another backend or decide to process the request itself. In our architecture the costly handoff operation si distributed among the backends which increase the scalability of the cluster. To test the efficacy of our proposed scheme, we performed trace driven simulations and compare the cluster throughput and cluster scalability of our scheme with scalability and throughput of the scheme proposed in [1]. The results show that our scheme achieves a greater degree of scalability than the scheme in [1], without any loss of throughput. 1

Citations