An empírical evaluation of simulation techniques for parallel simulation of message passing networks

Abstract

n the field of computer design, simulation is an essential tool to validate and evaluate architectural proposals. Conventional simulation techniques, designed for their use in sequential computers, are too slow if the system to simulate is large or complex. The aim of this work is to search for techniques to accelerate simulations exploiting the parallelism available in current, commercial multicomputers, and to use these techniques to study a model of a message router. This router has been designed to constitute the communication infrastructure of a (hypothetical) massively parallel computer. Three parallel simulation techniques have been considered: synchronous, asynchronous-conservative and asynchronous-optimistic. These algorithms have been implemented in three multicomputers: a transputer-based Supernode, an Intel Paragon and a network of workstations. The influence that factors such as the characteristics of the simulated models, the organization of the simulators and the characteristics of the target multicomputers have in the performance of the simulations has been measured and characterized. It is concluded that optimistic parallel simulation techniques are not suitable for the considered kind of models, although they may provide good performance in other environments. A network of workstations is not the right platform for our experiments, because the communication demands of the parallel simulators surpass the abilities of local area networks—the granularity is too fine. Synchronous and conservative parallel simulation techniques perform very well in the Supernode and in the Paragon, specially if the model to simulate is complex or large—precisely the worst case for traditional, sequential simulators. This way, studies previously considered as unrealizable, due to their exceedingly high computational cost, can be performed in reasonable times. Additionally, the spectrum of possibilities of using multicomputers can be broadened to execute more than numeric applications.