Dynamic Process Management System Architecture for Computational Clusters

Abstract

The problem of eficient utilization of computational resources of clusters arises as its load and number of users grow. Tasks like the fair use of resources and load balancing are common and should be solved by the operation environment of cluster. Existing mechanisms that solve the mentioned problems work ¯ne as long as parallel programs are run on a ¯xed number of resources. Allocating and freeing resources dynamically can highly improve the performance of a parallel program as well as the e±ciency of using the cluster. In the following paper the system architecture is described, which supports dynamic resource allocation and process spawning, which is alternate to MPI-2 standard's dynamic process spawning mechanism. It is also introduced, how dynamic task/process spawning can improve the performance of the parallel program.