1.2 Thesis Statement and Contributions

In this dissertation, we show that the key to effective utilization of cluster resources relies on a set of performance parameters, which eloquently delineates the performance capabilities of the target communication system. Through this set of performance parameters - the communication model, we can analyze, predict, evaluate and explain on issues related to high-speed communication on commodity clusters.

This communication model is a performance understanding tool, which captures the performance characteristics of the target machine and facilitates performance analysis as well as algorithm design. The model is based on a resource-centric view of the data flow through the abstract machine during a communication event. Most of the performance parameters can directly map to events or operations found in traditional message passing programming model. In addition, each performance parameter is associated with a microbenchmark that measures experimentally its associated cost unit, e.g. time measurement in $\mu s$ or clock cycle, that this parameter spends or takes to support the particular performance feature.

This dissertation exposes performance issues related to high-speed communication on commodity clusters. The principal contributions of this thesis are:

• We introduce a realistic communication model that supports performance understanding of the cluster communication system. This communication model becomes the foundation of this thesis research, which is used as a tool for supporting performance analysis and algorithm design.
• We provide a set of benchmark methodologies to quantify all performance parameters of our communication model. This communication model and its associated microbenchmarks make up as a tool for performance evaluation and analysis. As we believe that to assist performance understanding, the definition of the performance parameter should accompany with the benchmarking methodology that measures it.
• To study the congestion problem in high-speed networking, we exploit how congestion problem could affect the final performance from an architectural viewpoint. We show that how the buffering architectures of the switch interact with the communication protocol, and dominate the behavior of our lightweight messaging system under congestive loss situation. Our analytical and experimental results show that under asymmetric traffic loads, the output-buffered mechanism is more susceptible to the congestion loss problem than the input-buffered mechanism. During the modeling exercise, we identify salient features that enhance our understanding on the packet loss problem, and therefore, are relevant for the design and analysis of efficiency communication schedules.
• To avoid congestion loss, we have introduced a global congestion control scheme, which is a proactive approach in handling congestion. This scheme makes use of resource information provided by our model to prevent oversubscribing the network and avoid congestion loss. Through a global windowing concept, this scheme works on by collaborating all participating nodes to monitor and regulate the traffic load, which effectively avoids congestion loss and maintains sufficient throughput to maximize the performance.
• Based on the architectural model of the cluster system, we have devised an efficient communication schedule for complete exchange operation. The spirit of this algorithm is the node contention-free schedule operated at the packet level without explicit synchronization operation. This algorithm, the Synchronous Shuffle Exchange, effectively utilizes the communication pipelines and achieves high-performance. We make use of our communication model to show that this algorithm is optimal. In the experimental evaluation, we demonstrate that this algorithm is realizable and efficient, as it can reach 97% of the available bandwidth on various hardware platforms.