Parallelization of BiCGMisR method with cache-cache elements preconditioning
Tipus de documentComunicació de congrés
Condicions d'accésAccés obert
We consider Krylov subspace methods for solving a linear system of equations on parallel computer with distributed memory. For speed-up of parallel computation, it is necessary to shorten the communication time among processors. However, in paral- lelized Krylov subspace methods, global synchronization points for inner products cause increment of communication time. Thus, we created the strategy for reduction of synchro- nization points of parallel Krylov subspace methods. We transform the computation of parameter βk to reduce the number of synchronization points of various Krylov subspace methods per one iteration. In this paper, we apply this strategy to three-term recurrence and propose parallel BiCGMisR method as the eﬀective solver suited to parallel computer with distributed memory. Furthermore, through several numerical experiments, we make clear that parallel BiCGMisR method outperforms other methods from the viewpoints of both elapsed time and speed-up on parallel computer with distributed memory.