关于矩阵乘法的文献中有多种算法可以扩展到 MPI 范式。例如:
> 1Dsystolic [1]
> 2D-systolic, Cannon’s algorithm [2];
> Fox’s algorithm [3];
> Berntsen’s algorithm [4];
> DNS algorithm [5].
如果忽略矩阵特性(稀疏等),它基本上会恢复数据在进程之间的分布方式,以最大限度地减少同步和负载不平衡(每个进程之间分配的工作量)。
在这个recent work你可以看到两种不同的数据分布方式以及它们之间的比较。
论文:
[1] Golub G.H and Van C.H L., “Matrix Computations.”,Johns Hopkins University Press, 1989.
[2] Whaley R. C., Petitet A., Dongarra J. J., “Automated empirical optimizations of software and the ATLAS project” Parallel Computing 27, 1.2 (2001), 3.35.
[3] Fox G. C., Otto S. W., and Hey A. J. G., “Matrix algorithms on a hypercube I:
Matrix multiplication”,Parallel Computing, vol. 4, pp. 17-31. 1987.
[4] Berntsen J.,“Communication efficient matrix multiplication on hypercubes, Parallel Computing”, vol. 12, pp. 335-342, 1989.
[5] Ranka S. and Sahni S., “Hypercube Algorithms for Image Processing and Pattern Recognition”, Springer- Verlag, New York, NY, 1990.