Design, Implementation and Evaluation of Scalable Resource Management System for Internet Servers

A great deal of research has been devoted to solving the problem of network congestion posed against the context of increasing Internet traffic. However, there has been little concern regarding improvements of the performance of Internet servers such as Web/Web proxy servers in spite of the projections that the performance bottleneck will shift from networks to endhosts. For example, Web servers or Web proxy servers must accommodate many TCP connections simultaneously and their server throughputs degrade when resource management schemes are not considered. In the previous work, we have proposed a scalable resource management scheme for Web servers in order to overcome above-mentioned problems, which has two major mechanisms: one is a dynamic allocation of send socket buffer, which assigns a proper size of send socket buffer to each TCP connection based on an ‘expected’ throughput estimated at a TCP sender host. The other is a memory-copy reduction scheme, which decreases the number of memory access operations in a data transmission at a TCP sender host. We have validated the effectiveness of our proposed scheme through some simulation and implementation experiments, and confirmed that our scheme can improve a server performance and fair and effective usage of the send socket buffer can be achieved. In the current Internet, however, many Web document transfer requests are transferred via Web proxy servers, which are usually prepared by ISPs (Internet Service Providers) for their customers. They must accommodate many TCP connections both upward TCP connections (from the proxy 1 server to Web servers) and downward TCP connections (from client hosts to the proxy server). Hence, a proxy server becomes a likely spot for bottlenecks to occur during Web document transfers, even when the bandwidth of the network and Web server performance are adequate. In this thesis, therefore, we propose a novel scalable resource management scheme for Web proxy servers. Our proposed scheme has two following mechanisms: one is a dynamic allocation of receive socket buffer for each TCP connection at a TCP receiver host, which assigns it based on its utilization monitored at the receiver host. Another is a connection management scheme that intentionally tries to close ‘idle’ persistent TCP connections, which transfer no data, in order to prevent newly arriving TCP connections from being rejected when the server resources becomes shorthanded. We validate the effectiveness of our proposed scheme through some simulation and implementation experiments. From these results, we conclusively confirm that our proposed scheme can improve the server performance by up to 50% at maximum, and document transfer delay perceived by client hosts can be decreased by up to about 30%.