This dissertation shows how a realistic model can help system designers and programmers to understand the performance characteristics of the underlying communication system. The organization of this thesis is as follows. Chapter 2 introduces our communication model in detail. We first describe the architectural model that our communication model is built on, as well as provide a breakdown view on how data flow through this architecture. Then we list out all the parameters of this communication model together with their associated cost formulae. The last section in Chapter 2 is focusing on reviewing other related models, and show how our work is distinguished from other works.
Chapter 3 provides a brief description on the Directed Point communication package [59], which is used as our tool to validate the thesis statement. Then we measure two of the Ethernet-based implementations of Directed Point and perform a systematic analysis on their performance. Through our modeling framework, we easily spot out the strength and weakness of these systems. In Chapter 4, we extend our performance studies from a point-to-point analysis to a highly congested communication pattern, the many-to-one collective operation. We focus on the congestion behavior of how the reliable transmission protocol performs under heavy congestive loss situation. The primary network architectural feature relevant to this study is the switch's buffering architecture, which is one of the performance parameters of our communication model. We conduct both experimental and analytical studies on two different buffering architectures, and investigate on how these buffering mechanisms impact on the resulting performance.
Chapter 5 and 6 highlight another capability of our communication model - algorithm design and analysis. In these chapters, we are focusing on another communication pattern, the many-to-many complete exchange operation. In Chapter 5, we discuss an efficient algorithm, the Synchronous Shuffle Exchange algorithm, for the complete exchange operation with our communication model. We make use of our model parameters to show that this algorithm is optimal on a theoretical non-blocking network. We also show that, in reality, the switch's buffering architecture may hamper the performance of our optimal algorithm, due to the demanding nature of this algorithm.
In Chapter 6, we extend our study on the complete exchange operation to another realistic LAN topology, the Hierarchical network [35]. And we show that there are architectural limitations on this topology. Again, by using the resource information provided by our communication model, we have designed a proactive congestion control scheme to augment the original synchronous shuffle exchange algorithm to work efficiently on this network topology. Chapter 7 gives a summary of this dissertation as well as discusses on directions for future work. Last but not the least, Appendix A contains the full description of all benchmark methodologies of our communication model.