Collaborative Research: SI2-SSI: ELSI-Infrastructure for Scalable Electronic Structure Theory

预测，即所谓的从头算电子结构计算，特别是基于Kohn-Sham密度泛函理论(DFT)的计算，现在是一种广泛使用的科学主力，几乎应用于所有科学，并且越来越多地应用于工程和工业。在材料科学中，它们使具有改进性能的新材料的计算性设计成为可能。在生物学或药理学研究中，它们提供了对大分子或药物功能的分子水平的洞察。在寻找新能源解决方案的过程中，他们提供了对新的太阳能电池设计、催化过程和许多其他方面的分子水平的见解。在许多应用和计算中的一个关键瓶颈是具有系统大小的所谓Kohn-Sham特征值问题的“立方标度墙”(即，如果模型大小增加到10倍，则工作量增加1000倍)。该项目将建立一个开放源码软件基础设施“ELSI”，它为最初三种互补的解决方案策略提供一个通用的、实用的接口，以缓解或克服与解决Kohn-Sham特征值问题相关的困难。ELSI将支持广泛的最终用户社区，以不同的代码为中心，通常具有独特的功能，将一组专门的科学家与特定的解决方案联系在一起，轻松地为他们共同面临的关键问题整合最先进的解决方案策略。通过提供这些有效的、可访问的解决方案策略，我们将开辟电子结构理论的主要领域，而基于DFT的预测方法目前并不适用。这将反过来为材料科学、化学和所有相关领域的新发展打开大门。支持ELSI的承诺来自一些最重要的电子结构开发人员社区，以及高性能计算领域的行业和政府领导者。因此，我们将创建一个强大的美国基础设施，它利用来自全球活跃社区的大量用户和开发人员基础来开发用于材料研究的DFT方法。ELSI将支持和增强三种最先进的方法，每种方法最适合于特定的问题范围：(I)ELPA(用于Petascale应用的本征函数)库，用于有效地、大规模地并行解决本征值问题的领先库(对于高达1000个原子的中小型问题)，(Ii)在最近的重新实现中的OMM(轨道最小化方法)，其通过专注于简化的、(Iii)PEXSI(极点展开和选择性反转)库，这是一种针对一般系统(针对原子数在1,000秒及以上的问题)的经验证的缩减标度(至多二次标度)解决方案。通过建立许多电子结构开发人员已经熟悉的风格的标准化接口，ELSI将使生产电子结构代码能够显著减少本征值问题的“标尺墙”。首先，ELSI将帮助他们有效地使用可用的最强大的计算平台。目标平台是当前的大规模并行计算机和多核体系结构、基于GPU的系统和未来的多核处理器。其次，该项目将对ELPA、OMM和PEXSI进行有针对性的方法改进，例如，在非常大的系统中更有效地使用矩阵稀疏性。对类似的计算体系结构和类似的方法改进的关注将导致这些方法之间的重大交叉和协同作用。

Predictive, so-called ab initio electronic structure calculations, particularly those based on the Kohn-Sham density functional theory (DFT) are now a widely used scientific workhorse with applications in virtually all sciences, and increasingly in engineering and industry. In materials science, they enable the computational ("in silico") design of new materials with improved properties. In biological or pharmacological research, they provide molecular-level insights into the function of macromolecules or drugs. In the search for new energy solutions, they give molecular-level insights into new solar cell designs, catalytic processes, and many others. A key bottleneck in many applications and calculations is the "cubic scaling wall" of the so-called Kohn-Sham eigenvalue problem with system size (i.e., the effort increases by a factor of 1,000 if the model size increases by a factor of 10). This project will establish an open source software infrastructure "ELSI" that offers a common, practical interface to initially three complementary solution strategies to alleviate or overcome the difficulty associated with solving the Kohn-Sham eigenvalue problem. ELSI will enable a broad range of end user communities, centered around different codes with, often, unique features that tie a specialized group of scientists to that particular solution, to easily incorporate state-of-the-art solution strategies for a key problem they all share. By providing these effective, accessible solution strategies, we will open up major areas for electronic structure theory where DFT based predictive methodologies are not applicable today. This will in turn open doors for new development in materials science, chemistry, and all related areas. Commitments to support ELSI exist from some of the most important electronic structure developer communities, as well as from industry and government leaders in high-performance computing. Thus, we will create a strong U.S. based infrastructure that leverages the large user and developer base from a globally active community developing DFT methods for materials research.ELSI will support and enhance three state-of-the-art approaches, each best suited for a specific problem range: (i) The ELPA (EigensoLvers for Petascale Applications) library, a leading library for efficient, massively parallel solution of eigenvalue problems (for small- and mid-sized problems up to several 1,000s of atoms), (ii) the OMM (Orbital Minimization Method) in a recent re-implementation, which circumvents the eigenvalue problem by focusing on a reduced, auxiliary problem (for systems in the several 1,000s of atoms range), and (iii) the PEXSI (Pole EXpansion and Selective Inversion) library, a proven reduced scaling (at most quadratic scaling) solution for general systems (for problems with 1,000s of atoms and beyond). By establishing standardized interfaces in a style already familiar to many electronic structure developers, ELSI will enable production electronic structure codes that use it to significantly reduce the "scaling wall" of the eigenvalue problem. First, ELSI will help them make efficient use of the most powerful computational platforms available. The target platforms are current massively parallel computers and multicore architectures, GPU based systems and future manycore processors. Second, the project will make targeted methodological improvements to ELPA, OMM, and PEXSI, e.g., a more effective use of matrix sparsity towards very large systems. The focus on similar computational architectures and similar methodological enhancements will lead to significant cross-fertilization and synergy between these approaches.