A Self-adaptive Workload Balancing Algorithm on GPU Clusters

With the wide application of GPU in High Performance Computing, more and more heterogeneous CPU+GPU clusters have been established in many fields. But with the comprehensive using of heterogeneous CPU+GPU clusters, workload balancing has become an important problem when the process nodes coordinate with each other, and the execution time of a program on imbalanced clusters resides on the slowest node. Although there are many strategies and algorithms that can solve the problem of workload balancing to some extent, they generally face the problem of high consumption of communication caused by the task migration. In order to make up for the existing deficiencies, this paper proposes a virtual task migration algorithm adapted to GPU clusters on CUDA platform. This algorithm uses virtual task migration to avoid actual data transmission between nodes, so the communication overhead is obviously decreased. At last, this paper performs an actual test using matrix multiplication to verify this algorithm. The experiment results show that compared with static task partitioning, the algorithm proposed in this paper can effectively achieve dynamic workload balancing and reduce the execution time of programs on GPU clusters, thus the algorithm can significantly improve program execution performance of GPU clusters on CUDA platform.