摘要
VASP是材料微观物性模拟的重要软件,并行性能优良,其中矩阵对角化在其运算过程中占据大量时间。在HPC作业调度下,VASP并行的进程分配表现单节点和跨节点的差异。通过对MPI并行的节点内和节点间进程、EDDIAG和RMM-DIIS运行时间分析,明确了VASP程序并行中,矩阵对角化的重要性和OpenMPI对矩阵对角化并行的作用。并编写了一个通讯优化的矩阵对角化程序,通过测试并行程序托比串行的马斯算法表现出强大的可扩展性,基于OpenMPI的并行程序能够极大提高对角化的并行效率,但其并行度、通讯代价与计算规模之间的关系需要进一步研究。
VASP is an important software for material microscopic physical properties simulation, with excellent parallel performance, among which matrix diagonalization takes up a lot of time in its calculation process. Under HPC job scheduling, the parallel process allocation of VASP shows the difference between the single node and across multiple nodes. Through the analysis of the intra-node and internode processes of MPI parallel, the running time of EDDIAG and RMM-DIIS, we clarified the im-portance of matrix diagonalization in VASP program parallelism and the role of OpenMPI on matrix diagonalization parallelism. Furthermore, we write a communication-optimized matrix diagonalization program. By test, the parallel program using the Thomas algorithm shows strong scalability than the serial algorithm. The OpenMPI-based parallel program can significantly improve diago-nalization’s parallel efficiency, but the relationship between its parallelism, communication cost, and computing scale needs further study.
出处
《计算机科学与应用》
2021年第1期84-94,共11页
Computer Science and Application
