高效并行扫描问题是调度问题的子集,调度问题是NP完全问题.针对输运问题的特点,如何按特定的计算次序调度本地网格单元,以保证最佳的计算与通信性能是一个难度很大的问题.文中设计了一种基于局部深度优先的优先级(PDFDS)算法,该算法具有局部性、通信量小、优先级队列好等特点.将PDFDS算法应用到求解二维粒子输运方程的程序中,与现有的调度算法相比,新算法具有更好的并行计算效果,对于大规模计算问题,可以扩展到1024个处理器,相对于64个处理器的并行效率达到了96%.
Parallel sweep scheduling problem belong to the precedence constrained scheduling problem, and is NP-complete. How to schedule the local grids according to the characteristic of transport problem and keep the well performance of computation and communication is a hard work. This paper presents a priority scheduling algorithm based on DFDS algorithm. The new algorithm has well locality and less communication, and creates well priority queue. The new parallel priority scheduling algorithm is used to solve the two dimension particle transport equations. Performance results show that the new algorithm has better parallel performance than many scheduling algorithms widely used now. For the large transport problem, parallel efficiency is 96% on 1024 processors relative to 64.