An Approach for Maximizing Performance on Heterogeneous Clusters of CPU and GPU