An execution-time estimation model for heterogeneous clusters

Summary form only given. Heterogeneous clusters are flexible and cost-effective, but entail intrinsic difficulties in optimization. Although it is simple to invoke multiple processes on fast processing elements (PEs) to alleviate load imbalance, the optimum process allocation is not so obvious. Communication time is another problem. It is sometimes better to exclude slow PEs to avoid performance degradation, but it is generally difficult to find the optimal PE configuration. In this study, the execution time is first modeled from the measurement results of various configurations. Then, the derived model is used to estimate the optimal PE configuration and process allocation. We implemented the models from HPL (high performance linpack benchmark) of N = 400-6400, and estimated the optimal configuration for N = 3200-9600. The execution time of the estimated optimal configuration was 0%-3.6% longer than the actual optimal configuration. The models derived from N = 1600-6400 were also constructed, and their errors were 0%-4.3%for N = 1600-9600.