Network and Load-Aware Resource Manager for MPI Programs

We present a resource broker for MPI jobs in a shared cluster, considering the current compute load and available network bandwidths. MPI programs are generally communication-intensive. Thus the current network availability between the compute nodes impacts performance. Many existing resource allocation techniques mostly consider static node attributes and some dynamic resource attributes. This does not lead to a good allocation in case of shared clusters because the network usage and system load vary. We developed a load and network-aware heuristic for resource allocation. We incorporated the current network state in our heuristic. It is able to reduce execution times by more than 38% on average as compared to the default allocation.