相场模拟的并行计算方法.docVIP

Results ~8X Strong-scaling and efficiency plot for 108 × 108 × 108 system 21 Conclusion ? HYPRE was replaced by in-house Complex Iterative Jacobi solver ? ? Lesson learned: – Need better collaborative framework for ECSS (e.g. wiki, code MPI parallelization with non-blocking revision control such as git, communications for domainsubversion decomposition – Estimating reasonable goal can be OpenMP task parallelism for multiple difficult Helmholtz equation – Custom solver may give you better A scalable PFC code (weak and strong flexibility, but has development cost scaling, a major improvement