The authors propose a software mechanism for improving and guaranteeing performance of a Web server. The primary focus is on avoiding server overload (or in the event of a server overload, achieving graceful performance degradation), and on meeting the response time and throughput requirements of the hosted sites and end users.

Utilization of the server is computed by periodical measurement of the request rate and delivered bandwidth of the server. Then the utilization is used to determine whether or not the server is capable of meeting the required throughput and response time of its clients. In the case of an overload the clients are chosen for degraded service level or the nominal service level depending on the client’s priority assignment. An extension to the basic scheme is to share the underutilized server capacity among multiple virtual servers. Performance of a multiprocessor server with and without the proposed scheme was used in the study using a synthetic workload.

Although the authors claim that this mechanism is independent of the workload, it would have been interesting to see how their workload compares to real workloads.

The paper is well organized, but some of the sections could have been shorter. The proposed mechanism is valuable, and can certainly be used in today’s server environment. The experiments show that this mechanism meets its objective, that is, guaranteeing quality of service and allowing graceful degradation under overload conditions. The authors address the shortcomings of their scheme, which warrants further study.