Extending collective operations with application semantics for improving multi-cluster performance

We identify two ways of increasing the performance of allreduce-style of collective operations in a multi-cluster with large WAN latencies: (i) hiding latency in system noise, and (ii) conditional-allreduce where knowledge about the application is used to reduce the number of WAN messages. In our multicluster, system noise was not large enough to hide the WAN latency. But, the latency could be hidden using conditional-allreduce, since on many iterations only cluster-local values were needed, and many of the values needed from other clusters were prefetched. A speedup of 2.4 was achieved for a microbenchmark. Prefetching introduced a small overhead in the cluster with the slowest hosts.