一种面向CPU-GPU异构系统的容错方法

摘要

近年来,为了缓解日益严重的功耗问题,异构并行体系结构已成为超级计算机发展的一个重要趋势.图形处理器(graphics processing unit,简称GPU)凭借其超高的计算性能和性能功耗比,作为一种高效的加速部件已被广泛应用于高性能计算领域.但是,GPU先天的可靠性缺陷势必加剧超级计算机的可靠性问题.目前,国际上关于CPU-GPU异构系统容错技术的研究工作主要将GPU从异构系统中独立出来,以每次调用为粒度对其进行容错处理.设计了一种面向CPU-GPU异构系统的Lazy容错方法,给出了基于编译指导命令的容错框架及其约束,并讨论了相关的编译实现和优化方法,最后通过实验验证了该方法的正确性.实验结果表明,与现有的容错方法相比,利用所设计的LazyFT容错方法对GPGPU(general purpose computation on graphics hardware)程序进行容错处理,可以明显降低容错代价.%In recent years, heterogeneous parallel architecture has become an important development trend of supercomputer because it mitigates the problem of increasingly high power consumption. As a high performance and power efficiency accelerator, GPU (graphics processing unit) has been extensively used in HPC (high performance computing) area. However, the inherent unreliability of the GPU hardware deteriorates the reliability of supercomputer. Presently, most research of FT (fault-tolerance) techniques for CPU-GPU heterogeneous system isolates the GPU from the system, and does FT work for it at the granularity of a single GPU invocation. This paper proposes a new Lazy FT method for CPU-GPU heterogeneous system, introduces a FT framework and its constraints based on directives, and demonstrates the validity of the Lazy FT method. The experimental results show that, compared with existing FT methods, the cost of LazyFT is very cheap.