Constructing execution-time estimation models from diverse processing elements of heterogeneous clusters

Heterogeneous cluster is a reasonable extension of conventional PC clusters, while it is a difficult target for optimization. Although load imbalance can be alleviated by invoking multiple processes on fast nodes (without modification of source code), the optimal process allocation is not obvious. The preceding studies reported that practical estimation is possible by constructing the execution-time estimation models from homogeneous sub-clusters. In this study, we propose a method to construct the models from diverse processing elements of a heterogeneous cluster, and present some preliminary evaluation results. The derived models were accurate enough to find optimal or sub-optimal allocations, while requiring less nodes for model construction.