论文标题
征服代码:大规模和线性缩放DFT
The CONQUEST code: large scale and linear scaling DFT
论文作者
论文摘要
Conquest是从一开始就设计的DFT代码,可以在大规模平行的平台上实现极大的计算,从而为基态实现精确和线性缩放求解器。它使用局部基集(伪原子轨道,PAO和系统收敛的B-Splines)和稀疏的矩阵存储和操作,以确保计算各个方面的位置。使用精确的对角度方法和完整的PAO基集,可以使用相对适度的资源(200-500个内核)对多达1,000个原子的系统进行建模,同时使用多站点支持功能(MSSF),允许计算多达10,000个原子具有相似资源。通过线性缩放,该代码基本上表现出完美的弱缩放(每个过程固定原子),并且已应用于100万原子,将缩放到近200,000个核心;它已经在K计算机和Fugaku以及其他计算机上运行。征服可以准确计算总能量,力和应力,并允许对两个离子和模拟细胞进行结构优化。 NVE,NVT和NPT集合内的分子动力学计算是可能的,并且具有精确的对角度缩放和线性缩放[6]。代码与LIBXC接口以实现LDA和GGA功能,以及正在开发的Metagga和Hybrid功能。可以使用半经验方法(DFT-D2/3,TS)和VDW-DF进行分散相互作用。可以使用Resta的方法来计算极化。
CONQUEST is a DFT code which was designed from the beginning to enable extremely large-scale calculations on massively parallel platforms, implementing both exact and linear scaling solvers for the ground state. It uses local basis sets (both pseudo-atomic orbitals, PAOs, and systematically convergent B-splines) and sparse matrix storage and operations to ensure locality in all aspects of the calculation. Using exact diagonalisation approaches and a full PAO basis set, systems of up to 1,000 atoms can be modelled with relatively modest resources (200-500 cores), while use of multi-site support functions (MSSF) enable calculations of up to 10,000 atoms with similar resources. With linear scaling, the code demonstrates essentially perfect weak scaling (fixed atoms per process), and has been applied to over 1,000,000 atoms, scaling to nearly 200,000 cores; it has been run on both the K computer and Fugaku, among other computers. CONQUEST calculates the total energy, forces and stresses exactly, and allows structural optimisation of both ions and simulation cell. Molecular dynamics calculations within the NVE, NVT and NPT ensembles are possible with both exact diagonalisation and linear scaling[6]. The code interfaces with LibXC to implement LDA and GGA functionals, with metaGGA and hybrid functionals under development. Dispersion interactions can be included using semi-empirical methods (DFT-D2/3, TS) and vdW-DF. The polarisation can be calculated using Resta's approach.