| Recent years have seen a dramatic growth in the use of Web-based services. During this time, the underlying hardware hosting these services has evolved significantly. In particular, there has been a shift towards deploying Web servers on systems with multiple processor cores that aim to improve performance by exploiting parallelism within applications. These trends motivate the need to revisit scalability in Web servers. This research focuses on two distinct but inter-related issues;(1) validation of Web server benchmarking exercises, and (2) techniques to more effectively utilize multi-core systems.;The second part of this research quantitatively studies the scalability of a Web server on two different multi-core server architectures. The study finds that contention for shared resources such as the memory subsystem and the inter-CPU communication channels can adversely impact scalability. In particular, when considering a mean response time threshold of 0.5 msec, the maximum sustainable throughput on an AMD Shanghai server increases only by a factor of 4.6 when the number of cores increases from 1 to 8. To address this problem, this research proposes simple Web server software configuration strategies. Results show that these strategies can improve scalability by up to 29%.;Since modern multi-core systems are designed to sustain massive loads, it is important to validate the Web workload generators used in performance studies. This thesis proposes a methodology to conduct such validation exercises and applies it on a popular Web benchmark. The study finds that when generating moderate to heavy loads, the response times reported by the workload generator were often grossly inaccurate. Furthermore the generated workloads were less realistic than expected, causing server scalability to be in- correctly estimated. Consequently, it is crucial to not ignore systematic validation of workload generators as is often done in practice. |
