An Improved Execution-Time Estimation Model for Heterogeneous Clusters

A heterogeneous cluster can incur the performance degradation caused by the load unbalance in executing the application for homogeneous cluster. The total execution time can be sometimes improved by neglecting some of the PEs because communication time is reduced. This study examines to invoke multiple processes on fast processing elements (PEs) to avoid load unbalance. In this study, the execution time of each PE is firstly modeled from measurement results. Then, the derived model is used to estimate the optimal PE configuration and process configuration. Model composition is also proposed to reduce parameter extraction time. The derived models yield sub-optimal configurations in most cases, where the errors are less than 20%.