A study on optimal task decomposition of networked parallel computing using PVM

A numerical study is performed to investigate the effect of task decomposition on networked parallel processes using Parallel Virtual Machine (PVM), In our study, a PVM program distributed over a network of workstations is used in solving a finite difference version of a one dimensional heat equation, where natural choice of PVM programming structure would be the master-slave paradigm with the aim of finding an optimal configuration resulting in least computing time including communication overhead among machines. Given a set of PVM tasks comprised of one master and five slave programs, it is found that there exists a pseudo optimal number of machines, which does not necessarily coincide with the number of tasks, that yields the best performance when the network is under a light usage. Increasing the number of machines beyond this optimal one does not improve computing performance since increase in communication overhead among the excess number of machines offsets the decrease in CPU time obtained by distributing the PVM tasks among these machines. However, when the network traffic is heavy, the results exhibit a more random characteristic that is explained by the random nature of data transfer time.