Parallelisation of PyHEADTAIL, a Collective Beam Dynamics Code for Particle Accelerator Physics

TL;DR

This paper demonstrates significant speedup in particle accelerator simulations by parallelising PyHEADTAIL's longitudinal tracking engine using CUDA, achieving up to 100x acceleration for large-scale simulations.

Contribution

It presents the first parallel CUDA implementation of PyHEADTAIL's tracking engine, significantly improving simulation speed for particle accelerator physics.

Findings

Speedup of up to 100 times with CUDA implementation

Parallelisation effectively utilizes 448 CUDA cores

GPU acceleration outweighs overhead for large macro-particle counts

Abstract

The longitudinal tracking engine of the particle accelerator simulation application PyHEADTAIL shows a heavy potential for parallelisation. For basic beam circulation, the tracking functionality with the leap-frog algorithm is extracted and compared between a sequential C and a concurrent CUDA C API implementation for 1 million revolutions. Including the sequential data I/O in both versions, a pure speedup of up to S = 100 is observed which is in the order of magnitude of what is expected from Amdahl's law. From O(100) macro-particles on the overhead of initialising the GPU CUDA device appears outweighed by the concurrent computations on the 448 available CUDA cores.