Implementation And Performance Analysis Of All-to-all Communication

The MPI_Allgather and MPI_Alltoall are the most popular algorithms of collective communication of MPI, and they play an important role in parallel computation of numerous numerical and scientific application, such as FFT and matrix operation. The parallel program is deeply depended on the collective communication functions in MPICH library. With the popular of Cluster computer for parallel computation , more and more processors get into the parallel computation. It is difficulty to program ,only using Point-to-Point algorithm, such as MPI_Send and MPI_Recv. Collective communication is important to the parallel program , and its performance plays a more important role to the performance of the whole parallel computation.There are three modes for collective communication. They are one-to-all(broadcast), all-to-one, and all-to-all communication. The all-to-all communication has more manners for communication than the other two, and the different manners have different effects on the performance of function. Some issues are analyzed, which related to the high performance implementation of collective communication in this paper. They are the communication locality property, data count and communication step, in a simple model of a multiport fully connected message-passing system with TCP/IP. A new algorithm of MPI_Allgather and a new MPI_Alltoall for medium sized message are presented in this paper.We evaluate the new algorithm in ethernet cluster, and found the new MPI_Allgather sometimes performs better for medium-size and long messages than other algorithms. The new MPI_Alltoall is the better than other algorithms for medium-size messages.