Iterative solution of general sparse linear systems on clusters of workstations
Gen-Ching Lo, Yousef El-Mabruk Saad · 1996
Solving sparse irregularly structured linear systems on parallel platforms poses several challenges. First, sparsity makes it difficult to exploit data locality and this is true for both distributed and shared memory environments. Second, it is difficult to find efficient ways to precondition the system. For example, preconditioning techniques that have a high degree of parallelism often lead to slower convergence than their sequential counterparts. Finally, a number of other `global' computational kernels such as inner products can outweigh any gains due to parallelism, and this is especially true on workstation clusters where latency times may be high. In this paper we discuss these issues and report on our experience with PSPARSLIB, an on-going project for building a library of parallel iterative sparse matrix solvers. 1 Introduction In the past few years, there has been a flurry of activity on the use of distributed memory computers to solve challenging scientific problems. Partic...