Highly optimized simulations on single- and multi-GPU systems of 3D Ising spin glass

TL;DR

This paper introduces a highly optimized GPU-based Monte Carlo simulation for the 3D Ising spin glass model, achieving significant performance improvements through novel implementation techniques and multi-GPU parallelization.

Contribution

The paper presents new optimized methods for neighbor access, pseudo-random number generation, and multi-GPU parallelization in Monte Carlo simulations of spin glasses.

Findings

Achieved ~3 psFlip on GTX Titan with MINSTD PRNG

Achieved ~5 psFlip on GTX Titan with MT19937 PRNG

Demonstrated efficient multi-GPU implementation using MPI and CUDA streams

Abstract

We present a highly optimized implementation of a Monte Carlo (MC) simulator for the three-dimensional Ising spin-glass model with bimodal disorder, i.e., the 3D Edwards-Anderson model running on CUDA enabled GPUs. Multi-GPU systems exchange data by means of the Message Passing Interface (MPI). The chosen MC dynamics is the classic Metropolis one, which is purely dissipative, since the aim was the study of the critical off-equilibrium relaxation of the system. We focused on the following issues: i) the implementation of efficient access patterns for nearest neighbours in a cubic stencil and for lagged-Fibonacci-like pseudo-Random Numbers Generators (PRNGs); ii) a novel implementation of the asynchronous multispin-coding Metropolis MC step allowing to store one spin per bit and iii) a multi-GPU version based on a combination of MPI and CUDA streams. We highlight how cubic stencils and PRNGs are two subjects of very general interest because of their widespread use in many simulation codes. Our code best performances ~3 and ~5 psFlip on a GTX Titan with our implementations of the MINSTD and MT19937 respectively.