单节点多GPU集群下HPL动态负载均衡优化  

Optimizing HPL Benchmark on Multi-GPU Clusters

作　　者：陈任之[1] 黄立波[1] 陈顼颢[1] 王志英[1] 

机构地区：[1]国防科学技术大学计算机学院,长沙410073

出　　处：《计算机科学》2013年第3期107-110,共4页Computer Science

基　　金：国家自然科学基金(61070037;61103016)资助

摘　　要：现有GPU加速的高性能Linpack基准测试程序(HPL)一般采用基于实际运算能力的动态负载均衡算法来实现。然而该算法在单节点多GPU的平台上表现不佳,其原因是单节点多GPU平台上单个GPU计算量小,并且GPU与CPU的总性能差距较大。为此,提出了经验指导的动态负载均衡算法以及多GPU自适应负载均衡算法,并且在单节点多GPU平台上进行了验证,结果显示,其比现有的基于NVIDIA费米GPU的HPL有6.3%的加速效果。The current high performance Linpack benchmark accelerated by GPU usually employs the performance- based dynamic load balancing algorism. However, this algorism does not perform well on multi-GPU cluster. The reason is that each GPU on this kind of node has a smaller calculating scale, and the gap of the total performance between GPU and CPU is larger. Therefore, this article proposed an experience-based dynamic load balancing algorism and a multi- GPU adaptive load balance algorithm. Besides, the article tested these two algorisms on multi-GPU cluster and it a- chieves a 6. 3% acceleration compared with the latest NVIDIA＇s HPL accelerated by GPU.

关 键 词：HPL GPU 动态负载均衡算法 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]