Improving the Efficiency of GPU Clusters