An analytic model for predicting the performance of distributed applications on multicore clusters

Abstract

Computationally demanding applications can benefit from distributed processing. Distributed processing offers immerse computational power from collections of autonomous systems which can provide up to thousands of processing cores. Over the years, a considerable amount of research has been put towards developing performance prediction models for distributed applications. Amdahl's law states that the speedup of any parallel program has an upper bound which is determined by the amount of time spent on the sequential fraction of the program, no matter how small and regardless of the number of processing nodes used. This paper discusses some of the short comings of this law in the current age. We propose a theoretical model for predicting the behavior of a distributed algorithm given the network restrictions of the cluster used. The paper focuses on the impact of latency and bandwidth which affect the cost of interprocessor communication and the number of processing nodes used to predict the performance. The model shows good accuracy in comparison to Amdahl's law.