面向国产异构系统的HPL异构协同设计

Orchestrating HPL between CPU and China accelerator

HPL是高性能计算广泛采用的Linpack测试软件包,传统HPL算法中,求解矩阵将以块为单位循环分布到所有处理器,由于国产加速器(China Accelerator)的底层矩阵乘接口仅支持定制接口,传统HPL算法已不适合CPU+China Accelerator异构系统,因此,必须基于定制接口完成矩阵分布细致划分与封装dPEM,以提供一个通用的HPL测试配置环境;同时,为了充分发挥国产异构系统的效率,设计了异构协同矩阵乘调度算法OA4MM,以提高国产异构系统的效率。实验验证了dPEM的有效性和OA4MM算法的高效性,OA4MM较传统的异构HPL调度算法性能提升近10%。

HPL is a Linpack benchmark package widely used in high performance computing test.Matrix is divided into sub-matrix and distributed into computing elements in traditional HPL algorithm.However,it is ineffective for China Accelerator because of a specified interface on matrix multiplication built in China Accelerator.Thus,dPEM(delicate Partition and Encapsulation on Matrix)is advised to expose a friendly testing configuration environment.Furthermore,we propose OA4 MM(Orchestrating Algorithm for Matrix multiplication)based on heterogeneous system composed of CPU and China Accelerator.Experimental results validate dPEM and OA4 MM on CPU + China Accelerator.OA4 MM can promote productivity up to 10%in comparison to heterogeneous HPL.