Research On OpenMP Program Compilation And Optimization Techniques For Domestic Processors

With the increasing of high performance computers and the increasingly complex structure,the utilization rate of high performance computer systems is very low,and the inability of parallel programs to fully utilize the computing power of high-performance computing systems have become increasingly prominent.The problem of parallel programming has always been a major challenge for the development of high-performance computers,and its effective resolution plays a crucial role in giving full play to the role of high-performance computers and improving the efficiency of high-performance computer systems.There are more and more computing resources provided,and the structure becomes more and more complicated.It is more and more difficult for programmers to manually write parallel programs.Therefore,the research of parallelization technology seems to be more and more important.This paper focuses on the performance characteristics of domestic processors made by the “Sunway·Taihu Light”,and studies the loop scheduling strategy and parallelization techniques of OpenMP programs.The main work and contributions are as follows:At the beginning,this paper starts with the loop scheduling strategy,and divides the scheduling strategy of DOALL cycle into two types according to the way of accessing memory,and summarizes the two types of scheduling strategies respectively,and analyzes the scheduling overhead and load balancing of different scheduling algorithms in depth.A detailed comparison of the advantages and disadvantages of different scheduling algorithms was conducted.Secondly,analyze the scheduling strategy of DOACROSS loop.According to the different dependence distances,the DOACROSS loop can be divided into two categories.This article separately studies two types of loops and compares the advantages and disadvantages of the two types of DOACROSS loops.Thirdly,the research on the automatic parallelization of loops has been sorted out.For the first time,the parallelism of the loop is divided into SIMD vector parallelism,DOALL parallelism,DOACROSS pipeline parallelism and TASK pipeline parallelism,and detailed analysis is carried out.This article tests the standard test set program to verify the correctness and validity of this view.The research in this paper has certain guiding significance for the development of automatic parallelization system and optimization of application programs.