Unified Access To Kohn-Sham DFT Solvers for Different Scales and HPC: The ELSI Project

ORAL

Abstract

One major bottleneck in Kohn-Sham DFT is the solution of the Kohn-Sham eigenvalue equation to obtain the density matrix, which limits achievable system size. In this talk, we present updates to the ELectronic Structure Infrastructure (ELSI), an open-source effort that facilitates access to efficient Kohn-Sham solvers via a unified, code-independent interface for different system size scales, system types, and HPC architectures. Features that will be presented include support for several solver libraries (ELPA, libOMM, PEXSI, SIPs, CheSS) and a flexible, massively-parallel density-matrix-based restart method. Comparative benchmarks up to ten thousands of atoms on leading HPC resources will be shown. Special attention will be paid to the performance and scaling benchmark set created for this work, which is designed to be applicable to other aspects of electronic structure theory. These benchmarks will be the foundation of a decision layer to automatically select the optimal solver for a given problem based on the performance profile of the underlying HPC architecture.

Presenters

  • William Huhn

    Department of Mechanical Engineering and Materials Science, Duke University

Authors

  • William Huhn

    Department of Mechanical Engineering and Materials Science, Duke University

  • Alberto Garcia

    Institut de Ciència de Materials de Barcelona (ICMAB-CSIC)

  • Luigi Genovese

    CEA Grenoble, INAC - CEA

  • Ville Havu

    COMP/Department of Applied Physics, Aalto University

  • Mathias Jacquelin

    Lawrence Berkeley National Laboratory

  • Weile Jia

    Lawrence Berkeley National Laboratory

  • Murat Keceli

    Chemical Sciences and Engineering Division, Argonne National Laboratory

  • Raul Laasner

    Department of Mechanical Engineering and Materials Science, Duke University

  • Yingzhou Li

    Department of Mathematics, Duke University

  • Lin Lin

    Department of Mathematics, University of California Berkeley

  • Jianfeng Lu

    Department of Mathematics, Duke University, Mathematics, Duke University

  • Stephan Mohr

    Barcelona Supercomputing Center

  • Nkwe Monama

    Centre for High Performance Computing, Council for Scientific and Industrial Research

  • Happy Sithole

    Centre for High Performance Computing, Council for Scientific and Industrial Research

  • Alvaro Vazquez-Mayagoitia

    ALCF, Argonne National Laboratory, Argonne Leadership Computing Facility, Argonne National Laboratory

  • Chao Yang

    Lawrence Berkeley National Laboratory, Computational Research Division, Lawrence Berkeley National Laboratory

  • Haizhao Yang

    Department of Mathematics, National University of Singapore, Mathematics, National University of Singapore

  • Victor Yu

    Department of Mechanical Engineering and Materials Science, Duke University

  • Volker Blum

    Duke University, Department of Mechanical Engineering and Materials Science, Duke University, Mechanical Engineering and Materials Science and Chemistry, Duke University