Estimating the Optimal Configuration of a Multi-Core Cluster: A Preliminary Study

While multi-core processors became crucial elements of high-performance computing, the clusters of multi-core processors are still difficult target for optimization, since the communication time differs significantly in intra-node communication and inter-node communication. It is thus very important to select the optimal configuration of multi-core clusters; i.e., to find the optimal number of processes and to find the optimal allocation of processes to nodes. This study first elucidates the multi-core specific issues by examining the effect of process allocation in a multi-core cluster, using divergent and convergent allocations of four benchmark programs. Then, the estimation of optimal configuration is attempted for divergent and convergent allocations, using execution-time estimation models for single-core clusters. Though the estimation errors were less than 20% in most cases, further improvement is expected by incorporating the enhancements for multi-core systems into estimation models.