Collaborative Research: SI2-SSI: Removing Bottlenecks in High Performance Computational Science

Stanford University, Stanford CA

Investigators

Abstract

Computational chemistry has become ubiquitous throughout the fields of science and engineering. Among the many uses of computational chemistry codes are to aid in the design of new materials with specific properties, to understand protein folding dynamics related to the human body, to understand the mechanisms of enzyme catalysis to produce biofuels, and to identify and characterize gaseous species in the atmosphere. Because computational chemistry simulations can require large amounts of computer, memory and disk space, one focus of this research is to reduce the demands on computer resources by exploiting methods that can efficiently fragment large molecules into smaller pieces and at the same time take advantage of computers that have many thousands of computer cores. The principal investigators are developers of several computational chemistry programs, each of which has unique functionalities that have taken multiple person-years to develop and implement. In order to minimize further development efforts and to maximize the utility of the unique features, a second key focus of this research will be to develop the ability of each program to access the unique features and data of the other programs. A major bottleneck in the effort to construct high performance computers with more and more compute cores is the power that is required to drive such systems. One way that the research team will address the power issue is to explore the utility of low power architectures, such as graphical processing units for computational chemistry calculations. All of the newly developed codes will be made available to the user community by web download at no cost. This project presents an integrated computational science approach to very high quality electronic structure and dynamics calculations that will (a) be broadly accessible to both the development and applications communities, (b) have the capability to address problems of great interest, such as the properties of liquids and solvent effects, and photochemical/photobiological dynamics with high accuracy, (c) provide interoperability and sustainability into the foreseeable future, and (d) solve bottlenecks including power consumption using accelerators. A primary goal is to provide to the broad community new, accurate approaches that may be easily used, with a clear path forward to further code improvement and development. As part of the proposed effort, in addition to the usual outlets of journal articles and public presentations, the investigators will organize workshops at prominent national meetings, so that the expertise and software developed will be available to as broad a group of users as possible. All of the developed codes will be available on the web for easy downloads. The proposed research will develop new paradigms for interoperability, with a focus on the highly popular program suites GAMESS, NWChem, PSI4 and the AIMS dynamics code. Also included will be a cloud-based client-server model, a common quantum chemistry driver and novel data management approaches. The integration of the codes will be accomplished by making use of the combined expertise of the PIs in developing interoperable methods and data interfaces in computational science. In addition, several new methods will be developed, including novel explicit (R12) correlation methods that will be integrated with the most accurate levels of theory: multi-reference and the most accurate and novel coupled cluster methods, as well as the derivation and implementation of analytic derivatives. Applicability to large molecular systems will be made feasible by drawing upon novel fragmentation methods that scale nearly perfectly to the petascale.