Abstract

Sparse matrix-vector multiplication is the kernel for many scientific computations. Parallelizing this operation requires the matrix to be divided among processors. This division is commonly phrased in terms of graph partitioning. Although this abstraction has proved to be very useful, it has significant flaws and limitations. The cost model implicit in this abstraction is only a weak approximation to the true cost of the parallel matrix-vector multiplication. And the graph model is unnecessarily restrictive. This paper will detail the shortcomings of the current paradigm and suggest directions for improvement and further research.

Citations

616	A fast and high quality multilevel scheme for partitioning irregular graphs - Karypis, Kumar - 1999
171	The Chaco User’s Guide Version 2.0 - Hendrickson, Leland - 1995
19	Mesh partitioning and loadbalancing for distributed memory parallel systems - Walshaw, Cross, et al. - 1998
11	Partitioning sparse rectangular matrices for parallel processing - Kolda - 1457
9	Mesh Partitioning for Implicit Computations via Domain Decomposition - Farhat, Maman, et al. - 1995
1	Parallel algorithms for nonsymmetric iterative solvers - Hendrickson, Kolda
1	SCOTCH 3.1 user's guide, Tech. Rep. 1137-96, Laboratoire Bordelais de Recherche en Informatique, Universite - Pellegrini - 1996