Comparing the Performance of MC/DC's on-GPU Event-based Processing Methods in Multigroup and Continuous-energy Problems

TL;DR

This paper compares two GPU-based event processing methods in the MC/DC neutron transport package, analyzing their performance on benchmark and continuous-energy problems to inform scalable HPC applications.

Contribution

It introduces and evaluates a new asynchronous event processing method for GPU-accelerated Monte Carlo neutron transport simulations.

Findings

Asynchronous method shows better early scaling on the C5G7 benchmark.

Performance varies with problem type and parameters in continuous-energy simulations.

The study provides insights into GPU-based Monte Carlo method scalability.

Abstract

Monte Carlo / Dynamic Code (MC/DC) is a portable Monte Carlo neutron transport package for rapid numerical methods exploration in heterogeneous and HPC contexts, developed under the auspices of the Center for Exascale Monte Carlo Neutron Transport (CEMeNT). To support execution on GPUs, MC/DC delegates resource and execution management to Harmonize (another CEMeNT software project). In this paper, we describe and compare the performance of the two methods that Harmonize currently provides: a stack-based method and a distributed, asynchronous method. As part of this investigation, we analyze the performance of both methods under the 3D C5G7 k-eigenvalue benchmark problem and a continuous-energy infinite pin cell problem, as run across 4 NVIDIA Tesla V100s. We find that the asynchronous method exhibits stronger early scaling compared to the stack-based method in the 3D C5G7 benchmark. We also found that the asynchronous method exhibits mixed performance relative to the stack-based method in the continuous-energy problem, depending upon tally resolution, particle count, and transport loop decomposition.