Efficient Shared Memory and RDMA Based Design for MPI_Allgather over InfiniBand

MPI_Allgather is an important collective operation which is used in applications such as matrix multiplication and in basic linear algebra operations. With the next generation systems going multi-core, the clusters deployed would enable a high process count per node. The traditional implementations of Allgather use two separate channels, namely network channel for communication across the nodes and shared memory channel for intra-node communication. An important drawback of this approach is the lack of sharing of communication buffers across these channels. This results in extra copying of data within a node yielding sub-optimal performance. This is true especially for a collective involving large number of processes with a high process density per node. In the approach proposed in the paper, we propose a solution which eliminates the extra copy costs by sharing the communication buffers for both intra and inter node communication. Further, we optimize the performance by allowing overlap of network operations with intra-node shared memory copies. On a 32, 2-way node cluster, we observe an improvement upto a factor of two for MPI_Allgather compared to the original implementation. Also, we observe overlap benefits upto 43% for 32x2 process configuration.