An Innovative GPU-Optimized Multiscale Code for High-Fidelity Simulation of Collective Effects in Electron Beams

Old Dominion University Research Foundation, Norfolk VA

Investigators

Abstract

This award will lead to development of an innovative computer code for high-fidelity simulation of electron beams, drawing on recent advances in parallel computer architectures and applied mathematics. When electron bunches traveling at nearly the speed of light are forced by accelerator magnets to traverse a curved path, they emit bright ultraviolet or x-ray radiation. This radiation traveling a straight path catches up with the electron beam and can severely degrade its experimental usefulness. The first step in mitigating these damaging effects is to develop a trustworthy computer code capable of simulating the physics of this beam self-interaction. This project will enable the development of such a code. For the society in general, this research is a step in developing ultra-bright light sources which are essential tools for discoveries and innovations in physical, biological, energy and medical sciences. For example, they allow scientists to analyze chemical reactions, molecular structures and provide ways to develop new drugs. By training young researchers, this project will increase the number of much-needed accelerator scientists in the nation and make considerable effort in recruiting qualified female and under-represented minority candidates. Direct simulation of coherent synchrotron radiation (CSR) in two and three dimensions is prohibitively costly in terms of efficiency and memory requirements. Consequently, the present CSR codes employ various approximations that are inadequate for resolving essential physics in many realistic situations. These situations where existing CSR codes fail will become commonplace as the design of next-generation light sources commences. This project will design, benchmark and disseminate to the community a fundamentally new, multiscale, particle-in-cell code for modeling CSR, optimized to run on a parallel computational platform consisting of graphical processing units (GPUs). The code will exploit advantages of casting the problem in wavelet basis: (i) ability to retain information about the dynamics over the hierarchy of scales; (ii) removal of numerical noise by thresholding of the wavelet coefficients; and (iii) compact representation of data and operators, resulting in significant reduction of the computational load. Using NVIDIA's CUDA framework, a number of new code's vital numerical algorithms will be designed to run on a parallel GPU platform, and the successful matching of the algorithms and architecture will lead to code's scalability. The new code will be freely available to the community to enable modeling of a number of different machines.