A scalable approach for parallelization of Particle In Cell (PIC) based plasma simulations on supercomputers

Particle-In-Cell (PIC) method for plasma simulation tracks particle phase space information using a particle-data structure while mediating collective interactions via a grid-data structure. Charge Deposition (CD) subroutine in PIC method accumulates contributions from particles stored in particle data structure and interpolates charge density onto grid data structure. In parallel PIC codes, the CD subroutine often becomes a bottleneck due to the frequent interactions between these two data structures, which hinders scalability and efficiency because of dependency issues. Conventional parallelization methods require generating private grids for each core to remove these dependencies, but this approach has scalability issues. We propose a flag-based approach that requires only four private grids (replicas of the global grid) per node for a distributed memory system, or just four private grids for a shared memory system. This method enhances the scalability and performance of CD while maintaining the fundamental data structure of conventional methods. The private grids in shared memory provide complete accessibility for all concurrent threads. Additional functions and flags link particles to threads, avoiding simultaneous access to particles in the same cell. Performance evaluations using a low-temperature E×B plasma simulation with 1000 cores demonstrate the scalability of the flag-based approach. The fixed overhead associated with this method ensures high scalability.