Abstract

In this thesis, we present the design and implementation of the Concordia Parallel Systems Simulator (CPSS), a simulator for wormhole-routed multicomputers. The ultimate purpose of the CPSS is to provide a parallel programming environment which allows users to study impacts of system and software factors on program performance and to locate performance bottlenecks in parallel programs. The major challenge in the design of the CPSS is to make a good tradeoff between the accuracy of simulation results and the feasibility of simulation time. The CPSS can accurately simulate a large range of regular topologies that represent the communication structures of most applications in scientific computations as well as the topologies of many large-scale wormhole-routed networks. Users are given the flexibility of changing communication and computation parameters as often as needed. This flexibility allows for thorough analyses of program performance under different sets of systems parameters, or on various multicomputer systems having different characteristics. The CPSS provides a rich, powerful and user-friendly set of correctness and performance debugging tools. Performance statistics at several levels of details are available for users to fine-tune their programs. To support efficient network communication, we also propose optimal program mappings specifically designed for wormhole-routed networks. Our theoretical work in program mapping on wormhole-routed networks aims at emphasizing the importance of good mappings on such networks. As wormhole routing has become more popular, network sizes are expanded, and communication overheads are reduced, good mappings are indispensable to ensure high performance of applications running on wormhole-routed networks.