KHARMA: Flexible, Portable Performance for GRMHD

TL;DR

KHARMA is a new open-source, high-performance, modular code for general-relativistic magnetohydrodynamic simulations that efficiently runs on CPUs and GPUs, supporting various physics extensions and adaptive mesh refinement.

Contribution

It introduces KHARMA, a flexible, portable, and extensible GRMHD simulation code built in C++ with Kokkos, enabling efficient multi-physics and multi-scale astrophysical simulations.

Findings

KHARMA achieves high computational efficiency on CPUs and GPUs.

The code is modular, allowing easy swapping of physical and numerical components.

Validation and performance tests demonstrate KHARMA's accuracy and speed.

Abstract

KHARMA (an acronym for "Kokkos-based High-Accuracy Relativistic Magnetohydrodynamics with Adaptive mesh refinement") is a new open-source code for conducting general-relativistic magnetohydrodynamic simulations in stationary spacetimes, primarily of accretion systems. It implements among other options the High-Accuracy Relativistic Magnetohydrodynamics (HARM) scheme, but is written from scratch in C++ with the Kokkos programming model in order to run efficiently on both CPUs and GPUs. In addition to being fast, KHARMA is written to be readable, modular, and extensible, separating functionality into "packages," representing, e.g., algorithmic components or physics extensions. Components of the core ideal GRMHD scheme can be swapped at runtime, and additional packages are included to simulate electron temperature evolution, viscous hydrodynamics, and for designing chained multi-scale "bridged" simulations. This chapter presents the computational environment and requirements for KHARMA, features and design which meet these requirements, and finally, validation and performance data.