第一性原理软件北京原子技术模拟工具包的重构进展

Progress of refactoring first principle package of Beijing Simulation Tool for Atom TEchnology

基于第一性原理的理论方法的研究,代表了材料计算、分子模拟等领域的科学高地,相应的第一性原理计算软件直接关系到该领域相关理论、算法的积累.本文汇报了我们在重构第一性原理计算模拟软件—北京原子技术模拟工具包(BSTATE)的一些最新进展.重构的核心思想是降低用户使用门槛、扩展软件适用范围、增加软件对于流行计算框架的支持.基于此思路,在BSTATE原有Makefile编译系统的基础上添加了CMake编译环境,并支持各种数学函数库的自动和交互式配置;通过在原有内置泛函基础上增添Libxc泛函库的支持,使BSTATE支持的泛函数量有了数量级上的增长;分析测试BSTATE在集群的并行特点,并以更新数学库接口(FFTW3、Cufftw)的形式提供对于流行异构框架的初步支持.

The development of first principle methods can represent the summit of the sciences in the material computing and molecular modeling,and the corresponding first principle software packages are closely related with the accumulation of theories and algorithms in this field.In this paper,we reported our recent progress in refactoring the first principle package BSTATE.The key points in the reconstruction are lowering the doorsill,extending the scope of application,as well as adjusting package to the popular computer hardware.And as such,we updated the Makefile system to the new CMake system,in which the GUI can be used and many math libraries can be configured automatically;we added the support for the Libxc library,in which a large quantity of density functionals are included;we updated the interface for supporting GPU,in order to support the heterogeneous computing system.After refactoring,the Makefile system of BSTATE can supply both the Makefile and CMake system,and the Fourier transform libraries such as FFTW2,FFTW3,and Cufftw,the math libraries such as Intel MKL library,Openblas,and the density functional library such as Libxc,can be automatically or manually assigned.The integration of FFTW3 can slightly prompt the calculating efficiency in Intel’s many integrated core(MIC)architecture,and the integration of Cufftw can supply the initial support for the graphics processing unit(GPU)architecture,respectively.The usage of Libxc library makes the BSTATE package has the capacity to use hundreds density functionals,and the usages of various functionals were demonstrated by calculating the density of states of GaAs compound.Beyond the integration of various libraries,the parallel performance of BSTATE was also investigated.It can be found that the Fourier transformation and the solving for the eigenvalue equations are the major contributions.Using the tuning and analysis utilities(TAU)tool,we found that the tasks can be well distributed in modern HPC clusters.It implied that the refactoring didn’t affect the parallel efficiency of original BSTATE package.In a following benchmark test of graphene fragments,one can found that the refactored BSTATE package showed the best performance,its FFTW3&Libxc version owns about 0–17%acceleration comparing to that of FFTW2 version.