论文部分内容阅读
在分布式内存多处理机DMM（distributedmemorymultiprocessor）系统中，不同处理机上运行的任务之间的通信开销仍然很大，有时甚至抵消了多处理机并行所带来的好处．为了使并行程序在DMM系统上能得以高效的执行，必须采用合理的调度技术将任务分配给处理机．文章首先分别给出了任务调度系统中的任务模型、处理机模型以及调度问题的形式化描述，然后在此基础上研究了任务调度中3个最重要的问题，即（l）如何顺序选择参与调度的任务，（2）如何选择路由，（3）如何分配任务给处理机．其中，路由选择是按存储转发寻径与虫蚀寻径两种不同的方式来讨论的．最后，根据上述3个问题的解决策略，构造了一个非完全互连同构系统上的静态任务调度算法．
In the distributed memory multiprocessor (DMM) system, the communication overhead between tasks running on different processors is still very large, sometimes offsetting the benefits of multiprocessor parallelism. In order for parallel programs to be efficiently executed on a DMM system, tasks must be allocated to the processors using reasonable scheduling techniques. The paper first gives the formal description of task model, processor model and scheduling problem in task scheduling system, and then studies the three most important problems in task scheduling on the basis of (1) how to select and participate sequentially Scheduling tasks, (2) how to choose routing, (3) how to assign tasks to processors. Among them, the routing is stored and transmitted routing and worm eclipse routing two different ways to discuss. Finally, according to the above three strategies, we construct a static task scheduling algorithm on a non-fully interconnected homogeneous system.