单节点多GPU集群下HPL动态负载均衡优化

现有CPU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HPI有6.3%的加速效果。

Optimizing HPL Benchmark on Multi-GPU Clusters

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 achieves a 6.3% acceleration compared with the latest NV)DIA's HPL accelerated by GPU.