Blurring the Computation-Communication Divide: Extraneous Memory Accesses and their Effects on MPI Intranode Communications

摘要

Modern MPI simulator frameworks assume the existence of a Computation-Communication Divide: thus, they model and simulate the computation and communication sections of an MPI Program separately. The assumption is actually sound for MPI processes that are situated in different nodes and communicate through a network medium such as Ethernet or Infiniband. For processes that are within a node however, the validity of the assumption is limited since the processes communicate using shared memory, which also figures in computation by storing the application and its associated data structures. In this work, the limits of the said assumption's validity were tested, and it is shown that Extraneous Memory Accesses(EMAs) by a compute section could significantly slow down the communication operations following it. Two general observations were made in the course of this work: first, more EMAs cause greater slowdown; and second, EMAs coming from the compute section of the processes containing the MPI_Recv are more detrimental to communication performance than those coming from processes containing MPI_Send.