Planned AlltoAllv: A Cluster Approach
Adrian Jackson, Stephen Booth · 2004
The MPI AlltoAllv operation is generally based on a large number of point-to-point messages being sent by all the processes in the communication. This report aims to discover whether this process can be optimised for a clustered architecture by collating the data from all the processors within a node into a shared memory segment and then sending all the data for one node using a single message. This planned method has the potential to decrease overheads by greatly reducing the number of messages sent, and therefore diminishing latency, of the AlltoAllv call. We compare the performance of the planned AlltoAllv with that of the standard MPI AlltoAllv on a number of clustered architectures, and nd that for a standard benchmark the planned AlltoAllv can perform considerably better than the MPI version, but for a more realistic benchmark the performance benet is not as great. 1