High performance web server systems: design, testing, evaluation

The main goal of this thesis is to propose and investigate novel Web cluster architectures for scalable and high-performance fruition of static and dynamic Web content. We show the system bottlenecks intervening at different scenarios through the adoption of several case studies, pointing out the possible pitfalls and mistakes that can be done during the benchmarking process. An important premise is in order: all considered architectures, algorithms, and mechanisms are fully compatible with existing Internet/Web standards and protocols, so that they can be experimented and immediately adopted. The main contributions are outlined below. • We propose a classification of new architectures for locally distributed Web server systems. The analysis examines and characterizes the proposed systems through an original taxonomy, evidencing the trade-offs of the different techniques and pointing out the directions to develop more effective approaches. • We characterize the requirements of performance evaluation tools for a locally distributed Web server system and define the main components of a perfomance evaluation tool oriented to high performance Web clusters: workload characterization, traffic generation, data collection and analysis. For each of these components, we identify the main issues and discuss the alternatives, and also evaluate pros and cons of existing freely distributable tools, both academic and commercial. • We propose the design, the implementation and a detailed performance comparison of three different locally distributed Web systems: ClubWeb-1w, ClubWeb2wa, ClubWeb-2wk. All three prototypes are based on commodity off the shelf components, running under the Linux operating system. • We propose and evaluate the performance of several content-aware dispatching algorithms at the front-end and at the middle layer of an emulated, multi-tier Web cluster. We compare the performance of several combinations of centralized and distributed dispatching algorithms working at the first and second layer, and using different levels of state information. We also analyze the efficiency and the limitations of the different solutions and the tradeoff among the alternatives with the aim of identifying the characteristics of centralized and distributed approaches and their impact on performance. • We present a detailed performance evaluation and bottleneck analysis of an existing prototype dynamic, multi-tier Web auction site. The analysis shows the main bottlenecks and proposes improvements for different scenarios.