任务运行时系统StarPU在国产异构系统的移植和优化研究OA

With the rapid development of computer technology,the demand for computer perform-ance in scientific research and engineering has grown rapidly.At the same time,the emergence of accel-erators(e.g.,GPUs)has promoted the development of task runtime systems.Various computational examples(e.g.,Stencil computation and LU decomposition)are key steps in many fields of computing,and their execution efficiency directly affects the speed of solving practical problems.However,the de-velopment of high-performance computing on heterogeneous architectures in domestic computing system are still lacking.This paper ports the StarPU runtime system based on the C86 system.StarPU is a task-based runtime system that supports multiple heterogeneous architectures and can optimize algo-rithm operations through task scheduling and data management.The default scheduling strategy of StarPU has shortcomings in load balancing.Based on the scheduling strategy of StarPU,this paper designs and implements a scheduling strategy that integrates the two ideas of DMDAS and WS.The experimental results show that this scheduling strategy successfully transplants the task-based runtime system StarPU to the C86 system.Compared with the default scheduling strategy,DMDAS-WS achieves better execution efficiency on multiple benchmarks.