A performance comparison: MPICH, message passing interface against Treadmarks, distributed shared memory.

摘要

We compare two popular methods of programming in parallel environments: distributed shared memory (DSM) and message passing interface (MPI). This review discusses the application programming interfaces they present and the results of three practical examples of applications written for each environment.;The DSM system reviewed is the Treadmarks system and the MPI system is Message Passing Interface Chicago (MPICH). The programs involved are the Embarrassingly Parallel and Integer Sort algorithms from the Numerical Aerospace Simulation (NAS) Parallel Benchmarks Suite and Successive Over Relaxation (SOR-NZ) (or Jacobi algorithm) with non-zero elements. Two different data sets were used for IS, (IS-Medium and IS-Large).;Treadmarks performed at approximately 75% of the speedup level for MPICH on Embarrassingly Parallel for between two and eight processors. For eight processors, IS-Large Treadmarks achieved only 45% of the speed up of MPICH. Treadmarks attained 66% of the performance of MPICH for SOR-NZ.;Treadmarks offers a simpler, cleaner programming interface that allows the programmer to focus on the problem at hand, while allowing for the minimization of the details of communication, location and timing of data transmission. While an easier environment to program in, Treadmarks typically does not perform as well as MPICH. This is due to better messaging control in MPICH, and because of separation of data and synchronization in Treadmarks, in addition to the problem of false sharing.