Compressible Flow Solvers and Uncertainty Analysis on Million-Core Architectures
Speaker: Ivan Bermejo-Moreno, NASA Ames Research Center
Presentation
Abstract
Scaling and performance studies will be presented for two compressible flow solvers developed at the Center for Turbulence Research: Hybrid, a structured-mesh, finite-difference code used for DNS and LES of shock/turbulence interaction; and CharLESx, an unstructured-mesh, finite-volume code used for LES in arbitrarily complex geometries, including combustion and wall models. The tests were performed on the IBM Blue Gene/Q (BG/Q) Sequoia and Vulcan clusters deployed at Lawrence Livermore National Laboratory, comprising a total of 1,966,080 cores when combined. Hyperthreading, available in BG/Q, and its impact on the performance of the codes will be discussed. Several applications to massively-parallel simulations performed with these codes will be presented. In addition, the use of the parallel capabilities of the IBM BG/Q to accelerate uncertainty quantification studies will be highlighted with a case of application that includes over 1,300 3D RANS simulations of the interaction of shock waves and turbulent boundary layers in a nearly-square duct geometry. These RANS simulations, performed using the JOE solver also developed at the Center for Turbulence Research, are intended to replicate and complement experiments carried out at Stanford University, by incorporating aleatory and epistemic sources of uncertainty into the analysis.
Biography