Project A8 (finished)
A8. Development of 3D numerical radiation transport algorithms for the PHOENIX code on emergent parallel computer architectures
Background and Motivation
The topic of this research project is the development of 3D numerical radiation transport algorithms for the PHOENIX code on emergent parallel computer architectures.
Developed at the Hamburg Observatory, PHOENIX is a tool for modelling stellar and planetary atmospheres. This includes finding a numerical solution to the radiative transfer equation via the operator splitting method (see Kalkofen) while simultaneously computing the corresponding level populations of the atoms and molecules. Care was taken to not only model the case of LTE (local thermodynamic equilibrium), but also the more complex case of non-LTE, where a solution of the rate equations is necessary to determine level populations.
Due to the complex modelling and the huge amount of data involved in the considered problems, scientific computing makes extended use of parallel computing, which means data as well as task parallelism. The two most popular recent approaches in data parallelism are GPU-enhanced machines as well as manycore architectures. Currently, both kinds of machines are widely used among different fields of applications, such as astrophysics or meteorology.
Aims and Objectives
The spread of parallel architectures makes it necessary to develop numerical algorithms especially suited for those machines. A challenge for an effective data parallelization is the need to rethink serial structures in the code. Only "translating" serial to parallel code will most likely not result in a significant rise of computation speed, whereas designing appropriate numerical algorithms might.
Furthermore, interoperability of the parallelized code between different systems is crucial. It is necessary to find an implementation of PHOENIX that can be run on as many different systems as possible, independent of their exact build or manufacturer. There are various promising ideas, among them the OpenCL programming language, whose core feature is interoperability between different systems by different manufacturers. In consequence, the parallelized PHOENIX code could be run on various systems without the need for major modifications.
Developing and implementing numerical algorithms for PHOENIX especially suited for use with parallel architectures will hopefully result in a significant speed up of computation time. Additionally it should meet the needs of different types of parallel machines users have access to. This would in the foreseeable future enable PHOENIX users to strive for even more complex models and to include even more physical aspects, leading again to more detailed results.
PhD student: Victoria Wichert
Victoria Wichert successfully defended her PhD in April 2019