Methods to Enable Quantum Monte Carlo Simulations of Complex Functional Materials with Large Atom Count on Accelerators

APS  ·  2024

Ye Luo;

HIGHLIGHT

For this reason, simulations are limited by the memory capacity of each accelerator. In this work, we employ two strategies to overcome this limitation in QMCPACK.

RELATED

Papers      Patents     Grants     Venues     Experts    Cite    Share    Save

ABSTRACT

As leadership supercomputers converge to hybrid architectures which include multiple accelerators with in a compute node, managing data locality for optimal performance becomes a great challenge for application developers. To reduce expensive data movement between host and accelerators and among accelerators, frequently accessed datasets are made resident on accelerators. For this reason, simulations are limited by the memory capacity of each accelerator. In this work, we employ two strategies to overcome this limitation in QMCPACK. First, we implement hybrid orbital representation for GPUs using OpenMP offload to reduce GPU resident memory for a given system size. In the second scheme, data sets and computation are distributed across multiple GPUs. With these new features, quantum Monte Carlo simulations of functional materials with over a thousand atoms can be performed on Exascale supercomputers.

 Paraphrase     Questionfy     Wikify     Simplify     Summarize