Optimizing Communication for Clusters of GPUs