Development of MC/DC: a performant, scalable, and portable Python-based Monte Carlo neutron transport code

TL;DR

This paper introduces MC/DC, a Python-based Monte Carlo neutron transport code designed for high performance, scalability, and portability, verified through initial tests and benchmark comparisons.

Contribution

It presents a novel Python-based MC transport code with improved performance and scalability, utilizing code-generation libraries and an innovative abstraction and compilation scheme.

Findings

MC/DC runs hundreds of times faster than pure Python implementation.

MC/DC achieves about 2.5 times slower performance than Shift but with similar parallel scaling.

A new challenge problem based on C5G7-TD benchmark demonstrates MC/DC's capabilities.

Abstract

We discuss the current development of MC/DC (Monte Carlo Dynamic Code). MC/DC is primarily designed to serve as an exploratory Python-based MC transport code. However, it seeks to offer improved performance, massive scalability, and backend portability by leveraging Python code-generation libraries and implementing an innovative abstraction strategy and compilation scheme. Here, we verify MC/DC capabilities and perform an initial performance assessment. We found that MC/DC can run hundreds of times faster than its pure Python mode and about 2.5 times slower, but with comparable parallel scaling, than the high-performance MC code Shift for simple problems. Finally, to further exercise MC/DC's time-dependent MC transport capabilities, we propose a challenge problem based on the C5G7-TD benchmark model.