非结构网格瞬态电磁场计算中的高效通信方法

Efficient communication method for transient electromagnetic field computation on unstructured grids

针对使用非结构网格的时域间断伽略金方法在并行计算时通信复杂的问题,提出一种适用于瞬态电磁场大规模并行计算的最小通信周期策略。将进程间点对点通信的拓扑结构映射为通信矩阵,利用通信缓冲区未满时非关联进程的通信互不干扰特性,对互不干扰的进程通信顺序进行排序,将每轮同时进行的通信记为同一个通信周期,重新填充通信矩阵。最小通信周期策略对初始通信矩阵不断递归地取各元素余子式,每次递归结束后即得到一个通信周期中同时进行通信的进程,将这些进程对应的元素排除在下一次递归之外,直到所有初始通信矩阵所有元素排序完毕。最小通信周期策略能够有效降低并行迭代计算过程中的总通信周期数,减少通信过程消耗的时间,从而提升算法的计算效率。与传统策略相比,最小通信策略的通信周期数缩减到3%,并行效率显著提高,计算时间减少。同时,使用该策略在国产超级计算机神威·太湖之光上以8000核组(8000进程,520000核心)进行计算时取得了约70.38%(10倍扩展)的并行效率。

A minimum communication cycle strategy for large-scale parallel computation of transient electromagnetic fields is presented to solve the communication complexity of the discontinuous Galerkin time domain method using unstructured grids in parallel computation.The topology of point-to-point communication between processors is mapped to a communication matrix.By using the non-interference feature of communication between non-associated processes when the communication buffer is not full,the non-interference processors communication sequence is sorted,the simultaneous communication in each roundtrip is recorded as the same communication cycle,and the communication matrix is refilled.The minimum communication cycle strategy recursively takes the remainder of each element for the initial communication matrix,and after each recursion,processors that communicate simultaneously in the communication cycle are obtained,excluding the corresponding elements of these processors from the next recursion until all elements of the initial communication matrix are sorted.Minimum communication cycle strategy can effectively reduce the total number of communication cycles in the parallel iterative computing process,reduce the time consumed by the communication processor,and thus improve the computational efficiency of the algorithm.Compared with traditional strategies,the number of communication cycles of the minimum communication cycle strategy is reduced to 3%,which significantly improves the parallel efficiency and reduces the computing time.At the same time,the parallel efficiency of 70.38%(tenfold expansion)is achieved when using this strategy for calculation with 8000 core groups(8000 processors,520000 cores)on the light of the Sunway Taihu-Light supercomputer.