多核架构的构成是包括云计算在内的多种并行模式的硬件基础,而早期数值分解的一些细粒度并行算法在这种新架构下的效果不够理想。文中基于有向无环图的思想,通过对超级节点分块,将数值分解拆成大量的小任务异步执行,减少了多核架构带宽的影响,提高了并行效率,充分利用了多核的计算能力。5个大规模系统的最优潮流计算结果表明,所提出的方法加速效果显著,具有广阔的应用前景。
Despite the fact that the formation of the CPU multi-core architecture is the physical basis of parallel computation including cloud computation,the early fine-grained parallel algorithms cannot work well enough as expected in the new architecture.Based on the directed acyclic graph,the supernodes are split into blocks.The factorization operation is split into many asynchronously executed small tasks to reduce the influence of bus bandwidth with full use of the floating-point operation capability of all the cores.Numerical simulations on five systems show that the proposed method is very promising for large-scale application because of its significant speedup.