Accelerating Large-Scale Beam Dynamics Simulations with GPUs

The Advanced Photon Source (APS) at Argonne National Laboratory (ANL) is a third-generation, high-brightness, 7 GeV storage-ring-based x-ray light source. At present, Argonne is pursuing an APS upgrade project known as the APS Renewal, aiming at improvements through upgrades of beamline optics, detectors, and end-station equipment. Intensive computation is a pivotal aspect of this accelerator modeling and optimization eort, and signicant acceleration of already-existing computational tools used for this work will allow to explore a wider range of parameters and produce optimal designs in less time and at lower cost. Particle accelerator code ELEGANT has been used extensively for the APS upgrade work, as well as for numerous other light source design and optimization projects in the U.S. and abroad. We propose to develop an optimized, benchmarked set of CUDA kernels and drivers for GPU-accelerated simulation of beam dynamics with ELEGANT. We expect that this will translate into more than an order of magnitude speedup of the most time consuming simulations. We will closely collaborate with the ELEGANT code developers to ensure the optimal usage of the new capabilities. We prototyped a single-particle-dynamics kernel (quadrupole) and a collective-eects kernel (drift with longitudinal space charge), achieving a speedup of 80x and 36x respectively over a CPU implementation, and incorporated the GPU-accelerated algorithms into ELEGANT. We will develop and integrate into ELEGANT a comprehensive set of GPU-enabled algorithms; enable, automate, and optimize parallel computation on multiple GPUs; deliver enhancements to the ELEGANT software engineering and quality assurance process; and provide simulation support to the APS Renewal design optimization eort. Commercial Applications and Other Benets: Successful completion of this project will add to the functionality and value of the companys GPULib commercial library. In addition, the proposed GPU-accelerated beam dynamics modeling tools have the potential to be a signicant time and cost saver for both DOE-funded research institutions and the researchers in the Navy megawatt-class FEL program.