一种GPU集群的动态任务映射策略

A Dynamic Task Mapping Strategy of GPU Cluster

在3G网络的视屏计算中,需要解决逻辑复杂和数值运算密集的难题。为此,提出一种集中式动态任务映射策略。采用消息驱动方式,将任务初始分配,根据消息的内容进行实时优化分配,利用GPU集群的任务特征,设计对应的效益函数,以衡量任务映射策略的优劣。实验结果表明,该策略可以较好地协同各部分资源完成计算任务,并能适应集群的变化,在集群规模扩大或是有节点故障时均有比较稳定高效的性能。

The calculation of the screen in the 3G network, needs to solve the complex logic and numerical computation intensive problem, this paper proposes a centralized dynamic task mapping strategy, which uses message-driven approach to dynamically update the task queues. It can adapt to the scalability of clusters well. For the purpose of measuring the pros and cons of the strategy, this paper proposes a suitable benefit function according to the characteristics of the cluster＇s task. Experimental results show that this strategy can he well adapted to the scalahility of the cluster and have a good performance on practical task mapping with high revenue and throughput.