Performance analysis and modeling of regular applications on heterogeneous workstation networks

Heterogeneous networks of workstations and/or personal computers (NOW) are increasingly used as a powerful platform for the execution of parallel applications. Sometimes applications are developed having in mind this type of heterogeneous environment, but in most cases applications already developed for traditional parallel machines (homogeneous and dedicated) are ported to NOWs, resulting in performance degradation due in part to less efficient communications but more often to unbalancing. In this work we propose a simple model able to analyze and predict performance on heterogeneous NOWs of regular data-parallel applications originally developed for ring or 2-D mesh topologies. To improve performance, the computation time on the various nodes must be as balanced as possible. This can be obtained in two ways: by heterogeneous data partitioning or by assigning to each node a number of processes proportionally to its relative power. A test case based on matrix multiplication is analyzed and the results predicted by the model are compared with the ones collected experimentally. Our analysis shows that an efficient porting of homogeneous data-parallel applications on heterogeneous NOWs is possible and can be achieved in most cases in a quite straightforward and effective way.