Multi-GPU-based Swendsen-Wang multi-cluster algorithm for the simulation of two-dimensional q-state Potts model

TL;DR

This paper develops a multi-GPU implementation of the Swendsen-Wang multi-cluster algorithm for simulating the 2D q-state Potts model, achieving high performance and scalability on large-scale supercomputers.

Contribution

It extends previous single-GPU algorithms to a multi-GPU framework, enabling large-scale, high-speed simulations of the 2D Potts model.

Findings

Achieved 37.3 spin flips per nanosecond on 256 GPUs for the Ising model.

Demonstrated scalability and performance of the multi-GPU implementation.

Validated the algorithm's efficiency at the critical temperature for large system sizes.

Abstract

We present the multiple GPU computing with the common unified device architecture (CUDA) for the Swendsen-Wang multi-cluster algorithm of two-dimensional (2D) q-state Potts model. Extending our algorithm for single GPU computing [Comp. Phys. Comm. 183 (2012) 1155], we realize the GPU computation of the Swendsen-Wang multi-cluster algorithm for multiple GPUs. We implement our code on the large-scale open science supercomputer TSUBAME 2.0, and test the performance and the scalability of the simulation of the 2D Potts model. The performance on Tesla M2050 using 256 GPUs is obtained as 37.3 spin flips per a nano second for the q=2 Potts model (Ising model) at the critical temperature with the linear system size L=65536.