Abstract Details
Abstracts
Author: M Cianciosa
Requested Type: Poster
Submitted: 2019-02-22 20:15:13
Co-authors: M. Stoyanovmk, M. Beidler, D. Del-Castillo-Negrete
Contact Info:
Oak Ridge National Laboratory
1 Bethel Valley Road
OAK RIDGE, TN 37830
United States
Abstract Text:
Full orbit particle tracking is an easily parallelized problem allowing to simulate large numbers of particles using high performance computing (HPC) resources. For relativistic electrons generated during disruptions, timesteps on the order of ~1.0E-11 s are required to resolve the particle gyromotion. In order to reach the simulation time scales and particle statistics necessary for validation studies, include collisions, and for coupling to fluid and MHD solvers, efficient use of HPC resources is required. The Kinetic Orbit Runaway Code (KORC) simulates the full particle motion of large ensembles of energetic particles by integrating the relativistic Lorentz force. This work will present the results of CPU and GPU performance optimization. Using code profiling, bottlenecks in the layout of memory, implementations of interpolation routines for background plasma parameters, and random number generation for Monte Carlo collision operators were eliminated yielding significant performance improvements on Cori KNL nodes and other multicore CPU systems. Initial results and performance characteristics of a CUDA based implementation targeting the new generation of multi-GPU pre-exascale super computers such as SUMMIT will be compared against the CPU reference implementation.
Comments:
