Process Assignment in Multi-core Clusters Using Job Assignment Algorithm

Abstract

Modern high performance cluster systems for parallel processing are employing multi-core processors and high speed interconnection networks. Efficient mapping of the processes of a parallel application onto cores of such a cluster system, plays a vital role in improving the performance of that application. Parallel application can be modelled as a weighted graph showing the communication among the processes of that application. Such a graph can be constructed with the help of profiling tools. Cluster hardware also can be modelled as a graph, by collecting hardware details using HWLOC tool. Maximum weight matching based approach can be used to embed the application graph into cluster hardware graph. The proposed approach is implemented under a cluster system and tested using benchmark MPI parallel application. The performance of the parallel application, which is mapped using the proposed approach is better than, that is mapped using the legacy packed and round robin approaches of MPI library.