Modeling MPI Collective Communications on the AP3000 Multicomputer
The performance of the communication primitives of a parallel computer is critical for the overall system performance. The performance of the message-passing routines does not only depend on the hardware of the communication subsystem, but also on their implementation. When users migrate parallel codes to a multicomputer, they are unaware of the quality of the implementation and they can find that the performance of their parallel applications makes worse. In this work, we evaluate some representative collective communication routines of the MPI message-passing library: broadcast, scatter/gather, total exchange and reduction operations. Specifically, experimental results were conducted on the Fujitsu AP3000 multicomputer using Fujitsu MPI/AP version 1.0, in order to derive models for these MPI routines.
Unable to display preview. Download preview PDF.
- 1.Bailey, D., Harris, T., Saphir, W., van der Wijngaart, R., Woo, A., Yarrow, M.: The NAS Parallel Benchmarks 2.0. Tech. Report NAS-95-020, NASA Ames Research Center (1995) (Release 2.3 available at http://science.nas.nasa.gov/NAS/NPB)
- 2.Dongarra, J., Dunigan, T.: Message-passing Performance of Various Computers. Tech. Report UT-CS-95-299, Computer Science Dept., University of Tennessee, Knoxville (1995)Google Scholar
- 4.Ishihata, H., Horie, T., Shimizu, T.: Architecture for the AP1000 Highly Parallel Computer, Fujitsu Sci. Tech. J. 29(1) (1993) 6–14Google Scholar
- 5.Ishihata, H., Takahashi, M., Sato, H.: Hardware of AP3000 Scalar Parallel Server, Fujitsu Sci. Tech. J. 33(1) (1997) 24–30Google Scholar
- 7.Touriño, J., Doallo, R.: Performance Evaluation and Modeling of the Fujitsu AP3000 Message-Passing Libraries. To appear in EUROPAR’99, Toulouse, France (1999)Google Scholar