81. Exploring Performance of Domain Decomposition Strategies for Monte Carlo Radiation Transport
Authors: Platon Karpov (University of California, Santa Cruz)David Huff (New Mexico Institute of Mining and Technology)Xinyu Chen (University of New Mexico)Ryan Wollaeger (Los Alamos National Laboratory)Gabriel Rockefeller (Los Alamos National Laboratory)Brendan Krueger (Los Alamos National Laboratory)
Abstract: SuperNu is a Monte Carlo (MC) radiation transport code to simulate light curves of explosive outflows from supernovae. The MC transport step is domain replicated. To enable scaling on next-generation HPC systems, we implemented the recursive coordinate bisection approach of domain decomposition for the opacity calculation. Then, we propagated the decomposition to other steps in the simulation and constructed a communication infrastructure to support the decomposition. In this poster, we demonstrate the results of two communication schemes: the Improved KULL and Improved Milagro algorithms. We present results from a range of processors: Ivybridge, Haswell on local clusters at LANL, and the AMD Bulldozer Opteron 6200 series processors at Blue Waters’ Cray XE6 nodes. Finally, we tested the scalability of SuperNu on the latest Intel Xeon Phi architecture, Knights Landing.
Two-page extended abstract: pdf