面向神威众核超算系统的并行计算模型研究

Research on Parallel Computing Model for Sunway Many-Core Supercomputing System

基于异构众核处理器的超级计算机已经成为TOP500高性能计算机的主流,BSP、LogP、PRAM等已有并行计算模型均针对基于多核处理器的超级计算机设计,不能满足日益迫切的基于众核架构的超级计算机和应用发展需求.本文面向“神威·太湖之光”和神威E级原型系统的众核体系结构特点,提出P-PALN(Parallel-Parallel Access via LDM&NOC)并行计算模型,对于计算节点间的并行,该模型沿用BSP/LogP模型描述;对于计算节点内的众核并行,该模型提供私有存储访问和片上阵列通信的众核并行架构的有效描述PALN,能够协助用户进行众核并行算法设计,并在申威众核处理器硬件设计中指导参数的优化.实验结果表明,该模型可有效指导硬件设计和用户众核编程,从而提高系统和应用的性能.

Supercomputers based on heterogeneous many-core processors have become the mainstream of TOP500 high-performance computers.Existing parallel computing models,such as BSP,LogP and PRAM,are designed for supercomputers based on multi-core processors,which cannot meet the increasingly urgent needs of supercomputers and application development based on many-core architecture.Aiming at the Sunway Taihulight system and Sunway exa-scale prototype supercomputing system,based on the characteristics of many-core architecture,this paper proposes a hybrid parallel computing model P-PALN(Parallel-Parallel Access via LDM&NOC).For parallelism mode between computing nodes,the model uses the traditional BSP and LogP to describe.For parallelism mode within many-core computing node,this model effectively describes many-core parallel architecture featuring parallel private storage access to LDM and on-chip communication(PALN),which can guide users to better design parallel programming algorithms based on many-core architecture and engineers to optimize the corresponding parameters directionally during hardware design of Shenwei many-core processor.The experimental results show that the proposed model can effectively guide the hardware design and user many-core parallel programming,thereby improving the performance of the system and application.