Porting the particle-in-cell code OSIRIS to GPU-accelerated achritectures

Furthering our understanding of many of today's interesting problems in plasma physics requires large-scale kinetic simulations using particle-in-cell (PIC) codes. However, these simulations are extremely demanding, requiring that contemporary PIC codes be designed to efficiently use a new fleet of exascale computing architectures, which are increasingly GPU based. We discuss a GPU algorithm for PIC codes which we implemented on the code OSIRIS [1]. Development is currently ongoing, but a limited feature production code based on CUDA C is complete. Performance on GPUs is especially dependent on memory utilization, and maximizing utilization without exceeding capacity is challenging for many of the real-world problems where computational load can fluctuate dramatically in space and time. Our algorithm is unique compared to other PIC codes which have been ported to GPUs in that it includes two important features to overcome this challenge. First, it is being built on the existing OSIRIS data structures and will thus include dynamic load balancing. Second, we make use of a novel custom memory management strategy to avoid unnecessary—and costly—buffer reallocation in simulations where load imbalance is prevalent. We will also discuss performance and strategies for extending the software to run on non Nvidia GPUs.

References:

[1] Fonseca, R. A., et. al. (2002). OSIRIS: A Three-Dimensional, Fully Relativistic Particle in Cell Code for Modeling Plasma Based Accelerators. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2331 LNCS(PART 3), 342–351.

[2] Miller, K. G., Lee, R. P., Tableman, A., Helm, A., Fonseca, R. A., Decyk, V. K., & Mori, W. B.(2021). Dynamic load balancing with enhanced shared-memory parallelism for particle-in-cell codes. Computer Physics Communications, 259, 107633.