TY - JOUR

T1 - High performance collisional PIC plasma simulation with modern GPUs

AU - Romanenko, Alexey A.

AU - Snytnikov, Alexey V.

AU - Lazareva, Galina G.

PY - 2019/11/20

Y1 - 2019/11/20

N2 - GPU implementation of the collisional Particle-In-Cell plasma model is proposed. The goal of this model is the simulation of Plasma Enhanced Chemical Vapour Deposition (PECVD) reactors with glow discharge plasma and also for simulation of fusion plasma with sufficient role of collisions. Vlasov equation is being solved by the Particle-In-Cell (PIC) method. Collision are simulated with the null collision technique. The important difference from the collisionless approach is that major time is taken by collision evaluation and not by particle push. GPU performance is increased by storing the particles in cells and also by the exclusion of synchronization of threads at the stage of current and density evaluation. Electromagnetic field is given by either the Maxwell equations or be the Poisson equation for electrostatic case. The Maxwell equations are solved by the FDTD method. The Poisson equation is solved by the hybrid method. Performance obtained at the moment is 0.5 TFLOPS for particle push with the Nvidia Tesla V100, and the parallels efficiency is 92% for a cluster with 250 nodes with 2 GPUs each resulting in 500 GPUs.

AB - GPU implementation of the collisional Particle-In-Cell plasma model is proposed. The goal of this model is the simulation of Plasma Enhanced Chemical Vapour Deposition (PECVD) reactors with glow discharge plasma and also for simulation of fusion plasma with sufficient role of collisions. Vlasov equation is being solved by the Particle-In-Cell (PIC) method. Collision are simulated with the null collision technique. The important difference from the collisionless approach is that major time is taken by collision evaluation and not by particle push. GPU performance is increased by storing the particles in cells and also by the exclusion of synchronization of threads at the stage of current and density evaluation. Electromagnetic field is given by either the Maxwell equations or be the Poisson equation for electrostatic case. The Maxwell equations are solved by the FDTD method. The Poisson equation is solved by the hybrid method. Performance obtained at the moment is 0.5 TFLOPS for particle push with the Nvidia Tesla V100, and the parallels efficiency is 92% for a cluster with 250 nodes with 2 GPUs each resulting in 500 GPUs.

UR - http://www.scopus.com/inward/record.url?scp=85076232739&partnerID=8YFLogxK

U2 - 10.1088/1742-6596/1336/1/012005

DO - 10.1088/1742-6596/1336/1/012005

M3 - Conference article

AN - SCOPUS:85076232739

VL - 1336

JO - Journal of Physics: Conference Series

JF - Journal of Physics: Conference Series

SN - 1742-6588

IS - 1

M1 - 012005

T2 - 2nd Workshop on Numerical Modeling in MHD and Plasma Physics: Methods, Tools, and Outcomes, MHD-PP 2019

Y2 - 10 October 2019 through 11 October 2019

ER -