Performance tuning case study on graphics processing unit-accelerated monte carlo simulations for proton therapy

Abstract

Proton radiation therapy is one of the most effective modern methods of cancer treatment. Because radiation cannot distinguish between tumors and healthy tissue, simulating beams to predict the range of radiation is crucial. Monte Carlo simulation is the most accurate dose calculation method for radiotherapy. However, high accuracy requires extremely long computation time, thus limiting its clinical applications. To date, considerable efforts have been made to utilize graphics processing unit (GPU)-accelerated algorithm designs for accelerating the proton dose simulation. However, to completely utilize the capability of a specific GPU, the GPU configurations must be fine-tuned carefully. In this study, we propose the performance evaluation of GPU-accelerated Monte Carlo simulation programs for proton therapy. Considerable efforts have been made to properly use an autotuning tool to determine optimal configurations for the aforementioned programs. Furthermore, the source of counterintuitive observations was examined. Evaluation results show that appropriate configurations can considerably reduce the simulation time of programs.