Accelerating Large-Scale Excited-State GW Calculations on Leadership Class HPC Systems

 · Invited

Abstract

Large-scale GW calculations are state-of-the-art to accurately describe excited state phenomena in materials, which is critical for the design of novel new devices based on complex materials with applications in many fields. However, application of the GW method to complex systems is often perceived as being limited due to the high computational cost. Reduced time to solution can be achieved through novel methods, algorithms and optimal implementations on modern HPC systems. In particular accelerators such as GPUs can speed-up by order of magnitude conventional CPU-only implementations, and additionally reduce the energy per flop consumption. This talk showcases the various techniques used to achieve performance portability for the Material Science code BerkeleyGW on hybrid architectures targeting to accelerate large scale simulations with thousands of atoms. These techniques include the efficient use of accelerated libraries; asynchronous memory transfer, execution and overlap with MPI communication; batched operations; shared memory; and exploitation of high-performance memory to accelerate I/O. We achieve excellent strong and weak scaling on thousands of GPUs, and an order of magnitude or more reduction in time to solution compared to the CPU-only implementation. We demonstrate the scale of GW calculations to the order of over 10,000 electrons utilizing the entire Summit at OLCF (more than 27k GPUs) achieving over 100 PFLOP/s of double-precision performance and time to solution of the order of minutes.

*Resources for this work provided by NERSC, supported by the US Department of Energy (DOE) Office of Science under contract DE-AC02-05CH11231, and OLCF through the INCITE program, supported by the US DOE Office of Science under Contract No. DE-AC05-00OR22725. This work was supported by the Center for Computational Study of Excited-State Phenomena in Energy Materials (C2SEPEM), funded by the US DOE Office of Science under Contract No. DEAC02-05CH11231.

–

Presenters

  • Mauro Del Ben

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

Authors

  • Mauro Del Ben

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

  • Charlene Yang

    • NERSC, Lawrence Berkeley National Laboratory

  • Zhenglu Li

    • Department of Physics, University of California, Berkeley
    • UC Berkeley & Lawrence Berkeley National Laboratory
    • University of California at Berkeley, and Lawrence Berkeley National Laboratory
    • University of California at Berkeley and Lawrence Berkeley National Laboratory
    • Lawrence Berkeley National Laboratory
    • University of California, Berkeley
    • Lawrence Berkeley National Laboratory and University of California at Berkeley
    • Department of physics, University of California at Berkeley, Berkeley, California

  • Felipe Da Jornada

    • Stanford University
    • Materials Science and Engineering, Stanford University
    • Department of Materials Science and Engineering, Stanford University
    • Stanford Univ
    • Department of Materials Science and Engineering, Stanford University, Stanford, California

  • Steven G Louie

    • University of California, Berkeley
    • Department of Physics, University of California, Berkeley
    • University of California at Berkeley and Lawrence Berkeley National Laboratory
    • UC Berkeley & Lawrence Berkeley National Laboratory
    • University of California at Berkeley, and Lawrence Berkeley National Laboratory
    • Lawrence Berkeley National Laboratory
    • Department of physics, University of California at Berkeley, Berkeley, California

  • Jack Richard Deslippe

    • Lawrence Berkeley National Laboratory
    • NERSC, Lawrence Berkeley National Laboratory