APS Logo

BerkeleyGW on the Path to Exascale

ORAL

Abstract

On the path to exascale, graphics processing units (GPUs) dominate the HPC landscape, requiring developers to adapt and redesign their core implementations to effectively leverage the novel hardware paradigms. Here we present the various strategies employed to tackle these portability efforts for the BerkeleyGW software package. BerkeleyGW is a materials science software package employed to study the excited state properties of electrons in materials by using the GW and the GW plus Bethe-Salpeter Equation (GW-BSE) methods, and beyond. It is capable of scaling to hundreds of thousands of CPU cores and effectively utilizes strong scaling to tens of thousands of GPUs. We achieved over 100 PFLOP/s of sustained performance and over 50% of peak of Summit@OLCF, making GW calculations for thousands of atoms systems feasible within minutes. We discuss our software design to achieve true performance portability across various GPU vendor architectures by analyzing the performance of vendor specific programming models (CUDA and HIP) and the directives based portable counterparts (OpenACC and OpenMP-target). We highlight the challenges we encountered as well as the practices we found useful in the porting pipeline.

–

Presenters

  • Mauro Del Ben

    Lawrence Berkeley National Laboratory

Authors

  • Mauro Del Ben

    Lawrence Berkeley National Laboratory

  • Steven G Louie

    University of California at Berkeley, University of California, Berkeley, University Of California, Berkeley, University of California at Berkeley and Lawrence Berkeley National Lab, University of California at Berkeley; Lawrence Berkeley National Lab, UC Berkeley, University of California at Berkeley, and Lawrence Berkeley National Laboratory, UC berkeley, Lawrence Berkeley National Laboratory, Physics Department, UC Berkeley and Lawrence Berkeley National Lab, Department of Physics, University of California at Berkeley and Materials Sciences Division, Lawrence Berkeley National Laboratory, LBNL & UC Berkeley

  • Jack Deslippe

    NERSC