Slave-Core Parallel Implementation of Two-Dimensional Coupled Markov Chains on Shenwei Processors
Huaiyun Peng, Junyu Ye, Jiajing Wu, Luoyou Shan, Ziliang Wei, Xiaomei Zhao, Jingpeng Han, WU Shu-qin · 2024
The SW26010 pro processor on the new-generation Shenwei supercomputer is fundamentally different from ordinary commercial processors in that it adopts a master-slave heterogeneous multi-core hardware architecture, and almost all of the processor's computational performance comes from the slave core arrays. Therefore, if we use only MPI for parallelization at the task level and a simple compile-and-run porting approach, where only the master cores in each core group are involved in the task and the slave cores are left idle, this makes the Shenwei processor indistinguishable from the average commercial processor. In this paper, we first further understand the tasks to be processed, then partition them, and finally optimize the parallelization strategy according to the task characteristics to further partition the tasks that are computed on the master cores, so that each slave core is involved in the task and the full computational performance of the processor is exploited.