Large-Scale Condensed Matter DFT Simulations: Performance and Capabilities of the CRYSTAL Code
Large-Scale Condensed Matter DFT Simulations: Performance and Capabilities of the CRYSTAL Code
复制标题
DOI：
10.1021/acs.jctc.7b00687
复制
发表时间：
2017-10-01
影响因子：
5.5
通讯作者：
Dovesi, R.
中科院分区：
文献类型：
--
作者：
Erba, A.;Baima, J.;Dovesi, R.
如今，高性能计算资源的有效利用对于将第一性原理理论方法的适用性扩展到描述大型的、逐渐更加现实的分子和凝聚态系统至关重要。这只能通过为计算中最耗时的步骤设计有效的并行化策略来实现，考虑到量子力学算法的通常复杂性，这需要一些努力，特别是如果并行化要扩展到所有属性，而不仅仅是代码的基本功能。在这篇文章中，性能和能力的大规模并行版本的CRYSTAL 17软件包的第一性原理计算固体进行了讨论。特别是，我们提出：（i）允许进一步改进代码可伸缩性的最新发展（最多32768个核心）;（ii）在运行数千个核心的计算时，对代码的缩放和内存需求进行定量分析（最多约14 000）的原子每细胞;（iii）的代码的高数值大小一致性的文件;和（iv）概述了最近的从头算研究的几个物理性质（结构，能量，电子，振动，光谱，热力学，弹性，压电，拓扑）的大型系统研究的代码。
Nowadays, the efficient exploitation of high-performance computing resources is crucial to extend the applicability of first-principles theoretical methods to the description of large, progressively more realistic molecular and condensed matter systems. This can be achieved only by devising effective parallelization strategies for the most time-consuming steps of a calculation, which requires some effort given the usual complexity of quantum-mechanical algorithms, particularly so if parallelization is to be extended to all properties and not just to the basic functionalities of the code. In this Article, the performance and capabilities of the massively parallel version of the CRYSTAL17 package for first-principles calculations on solids are discussed. In particular, we present: (i) recent developments allowing for a further improvement of the code scalability (up to 32 768 cores); (ii) a quantitative analysis of the scaling and memory requirements of the code when running calculations with several thousands (up to about 14 000) of atoms per cell; (iii) a documentation of the high numerical size consistency of the code; and (iv) an overview of recent ab initio studies of several physical properties (structural, energetic, electronic, vibrational, spectroscopic, thermodynamic, elastic, piezoelectric, topological) of large systems investigated with the code.