【摘要】 在多计算机系统中,通信对性能的影响是至关重要的。很明显,通信开销越小越好,但不论从硬件上还是算法上来看,既然并行,通信是避免不了的。硬件上避免通信对计算影响的唯一方法是支持通信与计算重叠。算法上往往受实际问题本身的限制,很难避免通信。既然通信不可避免,那么,通信为多少时,多计算机并行处理才是值得的呢?本文利用概率理论分析了通信与计算应保持的比例,讨论了并行算法的粒度划分以及并行处理效率与通信的关系。更多还原

【Abstract】 The performance of a multicomputer is greatly affected by communication overhead. Certainly, no communication is the best case. However, for parallel processing in a multicomputer, no communication is impossible. The only way we can do is to minimize communication overhead from both hardware and software. From hardware aspects, communication and computation can be made overlap. From software aspects, what we can do is to optimize algorithms. However, there is a strong limitation for minimizing communication overhead, because communication is usually application dependent and multicomputer dependent. Now that we can not avoid communication, under at most how much communication, an application is worthy of parallel processing. In this paper, we study the ratio of computation to communication that should be kept, the granularity that should be, and the relationship between efficiency and communication overhead by the means of probabilistic theory.更多还原