Research On Scalable High Performance Web Server Systems

Now, the growth of the Internet has introduced two significant changes for Web server systems. First, with the increasing bandwidth of network, more and more local processing has related to remote service. Thereby, the user-perceived speed of computing increasingly depends on the ability of remote servers, which requires the high performance of Web server systems. Second, popular Web sites should be capable of servicing millions of hits simultaneously when special events, for example, Olympic Games, happen. Average throughput of these systems is small, but peak throughput is very high, which requires scalability of Web server systems.The general solution for this issue is to update the hardware of Web servers. However, Web servers could not fully utilize the hardware resource and present the performance matching it, which results from the poor support of OS for Web applications and the architecture of software discomforting the OS calls applied by Web applications. Another well-known solution is to use Web cluster servers with the single system image. Generally, the load balancer, also called front end, in a Web cluster system intercepts the HTTP requests from clients to servers and schedules each request according to the rules of load balancing. But, the performance of this kind Web cluster is always very poor, and does not scale with the increase of the number of servers. Furthermore, the deficiency of the load balancer includes no support for the scalability of storage, sophisticated scheduling strategies, the transaction integrity, and differentiated services. The reason for all of these problems is that the front end can not support the request routing mechanisms and the strategies for request distribution based on content. By now, researchers on Web cluster system have started to work on them, but the routing mechanisms and the scheduling strategies they have brought forward prevent further improvement of the performance and scalability of a Web cluster.We believe that the performance of the Web cluster is subject to three main factors: the performance of a single Web server, the routing mechanism, and the scheduling strategy, the latter two of which decide the scalability of the system. Hence, the research scope of this thesis will focus on these three aspects. The content includes:â‘ quantificational analyzing of I/O strategies and concurrent strategies of present high performance Web servers, and proposing ASPED (Asynchronous Single Process Event Driven) architecture. By the tests through SepcWeb99, the performance of Jaguar, a server based on ASPED architecture is better than that of Flash, the server based on AMPED (Asymmetric Multi-Process Event Driven), by 27.1%.â‘¡the research on present request routing mechanisms based on content and proposing the HTTP Splicing technology. By Httperf test, the performance of the front end based on HTTP Splicing technology is higher than that of the front end based on present TCP Splicing technology by 3.5 times.â‘¢though HTTP Splicing improves the throughput of...