AbstractsComputer Science

Physics and Computer Architecture Informed Improvements to the Implicit Monte Carlo Method

by Alex Roberts Long




Institution: Texas A&M University
Department:
Year: 2016
Keywords: Monte Carlo; Thermal Radiative Transfer; High Performance Computing
Posted: 02/05/2017
Record ID: 2064133
Full text PDF: http://hdl.handle.net/1969.1/156848


Abstract

The Implicit Monte Carlo (IMC) method has been a standard method for thermal radiative transfer for the past 40 years. In this time, the hydrodynamics methods that are coupled to IMC have evolved and improved, as have the supercomputers used to run large simulations with IMC. Several modern hydrodynamics methods use unstructured non-orthogonal meshes and high-order spatial discretizations. The IMC method has been used primarily with simple Cartesian meshes and always has a ?rst order spatial discretization. Supercomputers are now made up of compute nodes that have a large number of cores. Current IMC parallel methods have signi?cant problems with load imbalance. To utilize many core systems, algorithms must move beyond simple spatial decomposition parallel algorithms. To make IMC better suited for large scale multiphysics simulations in high energy density physics, new spatial discretizations and parallel strategies are needed. Several modi?cations are made to the IMC method to facilitate running on node-centered, unstructured tetrahedral meshes. These modi?cations produce results that converge to the expected solution under mesh re?nement. A new ?nite element IMC method is also explored on these meshes, which o?er a simulation runtime bene?t but does not perform correctly in the di?usion limit. A parallel algorithm that utilizes on-node parallelism and respects memory hierarchies is studied. This method scales almost linearly when using physical cores on a node and bene?ts from multiple threads per core. A multicompute node algorithm for domain decomposed IMC that passes mesh data instead of particles is explored as a means to solve load balance issues. This method scales better than the particle passing method on highly scattering problems with short time steps. Advisors/Committee Members: McClarren, Ryan G (advisor), Morel, Jim E (committee member), Adams, Marvin L (committee member), Amato, Nancy M (committee member).