HILO: Quasi Diffusion Accelerated Monte Carlo on Hybrid Architectures
Abstract
The Boltzmann transport equation provides high fidelity simulation of a diverse range of kinetic systems. Classical methods to solve the equation are computationally and data intensive. Existing stochastic solutions to the Boltzmann equation map well to traditional large multi-core and many-node architectures but suffer performance degradations on graphics processing units (GPUs) due to heavy thread divergence. We present a a novel algorithm, Quasi-Diffusion Accelerated Monte Carlo (QDA-MC), which improves performance on heterogeneous CPU/GPU architectures.
An equally important aspect of this project is the joint development of QDA-MC through collaboration between the computational and computer science communities. This collaboration identified computational platforms and features that best suit the algorithm, and influenced algorithmic details which improve its computational efficiency. In addition to algorithm details and implementation results, we present the code optimizations and the design decisions that were critical to the co-design process.
License
This code is released under LA-CC-11-076. The license is BSD-ish with a "modifications must be indicated" clause. See http://github.com/losalamos/HILO/blob/master/LICENSE for the full text.
Documentation
A Los Alamos technical report (LA-UR 11-05596) has been written for the project and is available in this repository as a PDF file.
Authors
Student authors developed code while interns at Los Alamos during the Summer 2011 Co-Design School (http://codesign.lanl.gov)
Students
Mahesh Ravishankar [email protected]
Jeffrey Willert [email protected]
Paul Sathre [email protected]
Han Dong [email protected]
Michael Sullivan [email protected]
William Taitano [email protected]
Los Alamos Mentors
Tim Germann
Dana Knoll
Bryan Lally
Patrick McCormick
Allen McPherson
Scott Pakin