Accelerating phase-change heat conduction simulations on GPUs

Abstract

Numerical simulation of the phase-change heat conduction process plays an important role in numerous engineering applications such as metal and alloy production and casting. Because of the increasing demands on the efficiency of the computational models utilized in advanced process control and optimization, high-performance computational models are becoming a necessary building block of practical applications. Thus, in this paper, we focus on accelerating the phase-change heat conduction simulations on GPUs. In addition to taking the advantage of thousands of GPU cores, the current paper presents a set of optimization methods to improve the resource usage of GPUs. The proposed methods include thread rescheduling, global function design, and a data caching mechanism, which helps establish an appropriate mapping from the numerical algorithm to the GPU hardware feature, thus remarkably enhancing the solving efficiency. We have validated the accuracy of the accelerated model with the Stefan problem and evaluated its performance with a three-dimensional steel solidification problem. A performance boost has been demonstrated, where the optimized GPU solver achieves a 333.6x speed-up compared to the non-parallel CPU solver. The performance characteristic of the accelerated solver makes it a powerful and promising tool for practical industrial usage.