Performance of multicomputers using high-speed communication links

Abstract

This thesis presents the results of a simulation study of the performance of a message-passing multicomputer using high-speed point-to-point communication links. The multicomputer system consists of IBM RS/6000 machines linked by 220 megabits per second fiber-optic links. This system is simulated using RIOSIM, a fast, accurate, and flexible execution-driven parallel architecture simulator. An accurate timing profiler, simulating the superscalar capabilities of the RS/6000 at runtime, generates dynamically timing estimates for the instructions executed. The simulation results are validated against actual measurements on a two-processor system, using a variety of algorithms. Results show that the errors are typically around 8%. The validated model is used to study systems with more than two processors. Simulation results indicate that this setup is suitable for coarse-grained parallel algorithms, some of which show almost linear speedups. For fine-grained algorithms, the high overhead in message passing proves to be a serious bottleneck, resulting in less than linear speedups.