Multidimensional Network Performance Table of Contents

Anderson, James Robert M.S.E.E., Purdue University, May, 1997. Multidimensional Network Performance. Major Professor: Seth Abraham. The interconnection network in a parallel processing system can have a profound impact on the success of a system in terms of performance or cost. The ability to clearly determine the trade-offs between performance and cost during the design of a network is critical to the development of a system that meets performance and cost constraints. A stochastic analysis of multidimensional networks with unidirectional or bidirectional links between nodes is presented. The analysis allows the development of an accurate model for examining the performance and cost trade-offs of different network configurations. The model is validated through simulation, and does not rely on the simplifying assumptions of previous models. In addition, the model is valid for the hypercube network. Using the classic network design constraints of constant link width, pin-out, and bisection width, it is demonstrated that network performance is highly dependent on offered load. To emphasize this point, the constant maximum throughput constraint is introduced, which compares networks on the basis of an equal ability to handle a range of traffic load. This constraint is fundamentally different than previous constraints, which are based on a characterization of hardware implementation costs. Results under the new constraint are very different than the results under the three classic constraints. For unidirectional networks, the results clearly show that a low dimensional network, while offering the lowest message latency, must be significantly more expensive than a comparable high dimensional network and in some cases may be impractical to implement. For bidirectional networks, the low dimensional network offers the lowest message latency with only a modest increase in hardware over the high dimensional network.