Challenges Adapting CUDA PIC Codes to multiple GPUs

2015

A Particle-In-Cell code is a common particle simulation method often used to

simulate the behaviour of plasma. In this work, a parallel PIC

code is developed in CUDA, with a focus on how to adapt the method

for multiple GPUs. An electrostatic three dimensional PIC code is

developed, with an FFT-based solver using the cuFFT library.

Several issues related to parallelizing the PIC code are discussed, along

with the performance on one and two GPUs compared to the

CPU version we developed based on OpenMP and FFTW as a benchmark.

For most problem sizes, the application is found to be memory bound,

with the speed of the memory interface playing a larger role than

the double precision performance. Alternatives to some of the naive

solutions are discussed, with suggestions for how the simulation

could be implemented for a shared memory computer. Benchmarks were

performed on several GPU platforms including Nvidia Tesla K20 and GTX 980,

and the challenges of getting the implementations to simulate plasma

oscillations are discussed.

NTNU