Evaluating and modeling communication overhead of MPI primitives on the Meiko CS-2
The MPI (Message Passing Interface) is a standard communication library implemented on a large number of parallel computers. It is used for the development of portable parallel software. This paper presents, evaluates and compares the performance of the point-to-point and broadcast communication primitives of the MPI standard library on the Meiko CS-2 parallel machine. Furthermore, the paper proposes a benchmark model of MPI communications based on the size of messages exchanged and the number of involved processors. Finally, the MPI performance results on the CS-2 are compared with the performance of the Meiko Elan Widget library and the IBM SP2.
Unable to display preview. Download preview PDF.
- 3.J.J. Dongarra and T. Dunigan, “Message Passing Performance on Varius Computers”, Tech. Report UT/CS-95-229, Computer Science Dept., University of Tennessee, Knoxville, 1995. Available at http://www.cs.utk.edu/~library/1995.html.Google Scholar
- 6.C.E. Leiserson, “Fat-trees: Universal Networks for Hardware-Efficient Supercomputing”, IEEE Transactions on Computers, vol. C-34, no. 10, pp. 892–901, 1985.Google Scholar
- 7.J.Miguel, A. Arruabarrena, R. Beivide, and J.A. Gregorio, “Assessing the Performance of the New IBM SP2 Communication Subsystem”, IEEE Parallel and Distributed Technology, pp. 12–22, Winter 1996.Google Scholar
- 8.M. Snir, S. W. Otto, S. Huss-Lederman, D. W. Walker, and J. Dongarra, MPI: The Complete Reference, The MIT Press, 1996.Google Scholar
- 9.Z. Xu and K. Hwang, “Modeling Communication Overhead: MPI and MPL Performance on the IBM SP2”, IEEE Parallel and Distributed Technology, pp. 9–23, Spring 1996.Google Scholar