The Parallel Speedup Improving Based on Wave Field Forward Communication Latency Hiding Technology
Lei Xie, Changming Zhao, Xi Wu, Tao Wu, Min Huang, Dan Xiao, Mingqing Yao, Chaoyang Xia · 2018
The Finite Difference Method (FDM) is the most common parallel computing method for high performance numerical simulation. The method is able to solve and obtain physical field parameters efficiently when grid strengthening treatment is dense enough. However, FDM processing in parallel acceleration case is necessary to divide all raw data into several parts, which is corresponding to the number of computing nodes amount. It may cause that the edge raw data of each computing node are essential to communicate with the other edge data of the surrounding six adjacent computing nodes. Once the system communication time step is short enough and the computing node is of high computing performance, the communication latency may turn to the primary reason for inefficient parallel. In this paper, it provides a scheme for communication latency hiding that part edge data of each node is able to running in communication and computing simultaneously. The experimental results show that the proposed method has the ability to improve parallel efficiency and reduce the running time definitely.