Comparing architectures using throughput-versus-cost modeling

The paper compares two parallel architectures in terms of throughput versus cost. Throughput is estimated using machine parameters extracted from detailed models of each architecture. Cost models are used to express total resource use in terms of a common unit. Performance of an architecture is then evaluated by optimizing throughput for each possible cost. Finally, these throughput-versus-cost results for the two architectures are compared for a particular class of problems: iterative computations on a 2-dimensional grid. The results show that there is a cost below which one architecture is an order of magnitude faster than the other, and above which this relationship is reversed. This approach may prove useful as a general comparison methodology.<>