抄録
GPBiCG (Generalized Product of BiCG) method is often used for the solution of realistic problems on parallel computers. However, the number of synchronizations points for inner products needs three times per iteration, and these synchronization points cause great increase of communication time. Then, we devised a strategy to reduce the number of synchronization points. We applied this strategy to the original BiCGSafe method and succeeded in fair reduction of their communication time. In this paper, we apply this strategy to the original GPBiCG method. Through many numerical experiments, we will make clear that the proposed parallel version of GPBiCG method with double synchronization points outperforms the conventional methods in view of parallel computation.
