Predicting MPI buffer addresses.

Abstract

Communication latencies have been identified as one of the performance limiting factors of message passing applications in clusters of workstations/multiprocessors. On the receiver side, message-copying operations contribute to these communication latencies. Recently, prediction of MPI messages has been proposed as part of the design of a zero message-copying mechanism. Until now, prediction was only evaluated for the next message. Predicting only the next message, however, may not be enough for real implementations, since messages do not arrive in the same order as they are requested. In this paper, we explore long-term prediction of MPI messages for the design of a zero message-copying mechanism. To achieve long-term prediction we evaluate two prediction schemes, the first based on graphs, and the second based on periodicity detection. Our experiments indicate that with both prediction schemes the buffer addresses and message sizes of several future MPI messages (up to +10) can be predicted successfully.