A communication staging technique for network cache interconnected clusters

The work is an experimental investigation of a technique for staging communication between tasks of parallel applications in clusters that are interconnected with network cache architecture. With network cache, a process can write broadcast data to a shared buffer on the NIC. This data is immediately broadcast on the network and is available in shared buffers on the NICs of all workstations. By using network cache as a large staging area for collective communications, the required number of data and synchronization messages is reduced and application performance is improved. A version of matrix-vector multiply in which the result vector is available to all processes at the end of the computation is implemented, both using network cache and using traditional MPI messaging libraries. The implementations are compared experimentally and results show that completion time is improved by several factors with network cache staging rather than the point-to-point implementation of MPI messaging libraries.