Prediction of parallel algorithms performance on bus-based networks using PVM

Abstract

We adapt the classic sequential model used to predict the performance of communications on parallel computers to a local area network using PVM. We have discovered that the linear model using the values of the parameters (latency, transfer rate) given by the "ping-pong" algorithm does not predict correctly the times invested in several communication patterns. This work concentrates in one of these patterns: One to All communications. We propose an alternative experiment that takes into account the overlapping between computations and communications that arises for this pattern. We illustrate the problem and its solution using a parallel matrix multiplication algorithm based on the master-slave paradigm. The detection of this phenomenon and the determination of the new values far the parameters simplify the design of parallel algorithms and can be used to predict the appropriate number of processors to achieve maximum efficiency.