面向通用计算GPU集群的任务自动分配系统

Automatic Task Assignment System of General Computing Oriented GPU Cluster

当前GPU集群的主流编程模型是MPI与CUDA的松散耦合,采用这种编程模型进行编程,存在编程复杂度大、程序的可移植性差、执行效率低等问题。为此,提出一种面向通用计算GPU集群的任务自动分配系统StreamMAP。对编译器进行改造,以编译制导的方式提供集群任务的计算资源需求,通过运行时系统动态地发现、建立并维护系统资源拓扑,设计一种较为契合GPU集群应用特征的任务分配策略。实验结果表明,StreamMAP系统能降低集群应用程序的编程复杂度,使之较为高效地利用GPU集群的计算资源,且程序的可移植性和可扩展性也得到了保证。

MPI＋CUDA are the mainstream programming models of current GPU cluster architecture. However, by using such a low level programming model, programmers require detailed knowledge of the underlying architecture, which exerts a heavy burden. Besides, the program is less portability and inefficient. This paper proposes StreamMAP, an automatic task assignment system on GPU clusters. It provides powerful, yet concise language extension suitable to describe the compute resource demands of cluster tasks. It develops a run time system to maintain resource information, and supplies an automatic task assignment for GPU cluster. Experiments show that StreamMAP provides programmability, portability and scalability for GPU cluster application.