三阶段可任意划分负载应用的多次数据分配

A collection-aware multi-round scheduling algorithm

为了进一步提高数据并行应用在星型网络上的运行性能,基于多次数据分配策略提出了一种调度算法.在该算法中一个处理器将输入数据和计算结果分成多次接收和发送,在计算的同时接收下一阶段的输入数据和发送上一阶段的计算结果,使得计算和通信更好地进行重叠而最终降低了整个应用的响应时间.分析了系统各参数对算法的不同影响,实验表明:与广泛采用的LIFO和FIFO算法相比,本文提出的算法可以使用更少的资源而获得更好的调度质量.

To increase the efficiency of data-parallel application on the star network, this paper proposes a heuristic CAMR （Collection-Aware Multi-Round） for the collection-aware divisible load scheduling based on the multiround distribution strategy. It can overlap the computing and communication phases on the same conductor, thus decreases the response time of the whole application. We analyzed the impact of system parameters on CAMR. Experimental studies show that CAMR has better performance than the well-known LIFO and FIFO algorithms.