Radiation-hydrodynamics on massively parallel architectures

Abstract

This thesis presents a significant advancement in the capabilities of the radiative transfer package Taranis by coupling it to the widely used GIZMO hydrodynamics code. Together these two packages allow for fully coupled cosmological radiation-hydrodynamics simulations. The radiative transfer uses ray tracing and is performed entirely on graphics processing units (GPUs). It has been parallelized to be able to utilize multiple GPUs on the same shared memory node to significantly improve performance.

A significant amount of work went into optimizing Taranis to allow it to handle the vastly increased source count found in cosmological simulations as well as addressing some of the pitfalls found in previous versions of the code. In cosmological test cases improvements of a factor of > 10³ have been observed. Taranis also performs well in hydrodynamic test cases when compared to similar codes.

In cosmological settings, Taranis generates realistic sources on the fly based either on star particles or halos. These were used to run a set of cosmological simulations to investigate the beginning of the Era of Reionization. The spectral shape was found to have a significant impact on the pre-heating of the IGM ahead of the ionization fronts. Gas is rapidly heated to ~ 10⁴ K with mostly low-density gas being ionized first.