EPE00002

Performance and GPU compatibility

Author: Shreyas Bapat

date-created: 2020 February 02

date-last-revised: 2020 February 02

date-accepted: 2020 February 02

type: Standard Track

status: Discussion

Abstract

The EinsteinPy Codebase is currently in a beta state, where it can be used to play around with the data but, lacks the possibility of serious computations due to bad performance. We recently made some performance enhancements to accelerate the code upto 10-15X the current speed. But it is still not enough. Serious computations and research requires a very good performance along with ability to scale on multiple CPUs, GPUs. This can further facilitate the adoption of EinsteinPy in general relativity community. Moreover, the overall code structure, including coordinate conversions are little bit tricky as well as adding a new frame/coordinate system is a tiresome process in current model.

Detailed description

There are multiple ways of accelerating the codebase. We would like to slowly try each and every one of them. First step in determining the performance will be profiling the complete codebase using asv!

After profiling is done, we can have two backend type model (like tensorflow [Also, the tensorflow installation is now quite easy using conda]). Or anything else. For the low performance codebase too, we will need to replace certain Python snippets with either less comples pyton jitted code snippets wherever possible, or with some sort of compiled fast language (C++ / Rust).

The coordinate conversions in EinsteinPy are not the best in the industry right now. We aim to fix the problem of coordinate conversion through this EPE.

Some modules like, shadow do not scale well. Also there are some theoretical issues which make the whole module slow. We would like to simulate black hole shadows of real sized black holes rather than M=1 Black holes which have no practical significance.

Finally, this EPE also covers the performance of the visualisation module. There is a need for unification of plotting. There are 4-5 different types of plotters available in the codebase which majorly confuse users.