NYCCS/Mechanical Engineering Seminar

Hae-Won Choi

 University of Colorado, Boulder
 
 “High Performance Computing Technologies Devising
High-Order Methods and Adaptive Mesh Refinements for Computational Fluid Dynamics”
 
Thursday, March 1, 2007 @ 10:30 AM
Light Engineering Bldg., Rm. 250

Abstract: Recent paradigm in large-scale scientific computing have motivated investigations into computational fluid dynamics (CFD) and numerical methods that are more suitable for distributed-memory parallel computers. The future evolution of high-performance computing (HPC) will require a highly scalable, accurate and efficient computational algorithm. Accurate numerical schemes are essential to ensure high-fidelity simulations capable of capturing the multi-physical, multi-scale, and complex phenomena. To construct efficient computational meshes for a variety of CFD and numerical methods, adaptive mesh refinement (AMR) with suitable error estimators needs to be devised. Scalable performance is necessary to exploit the massively parallel petascale systems that will dominate HPC for the foreseeable future. The goal of my thrust will be research into implementation of advanced computational technologies through high-order methods devised with a scalable parallel computing algorithm.  The applicability and efficiency of CFD and numerical methods span a wide variety of scientific and engineering applications. Despite the relative maturity and widespread success of CFD and numerical techniques in applied science and engineering, there remain open challenges that need to exploit the full potential of current high-performance parallel computers. In order to address some of open issues, my talk will focus on some of recent efforts to achieve the goal of my future research. First, I will address parallel computations for emerging engineering applications: the electro-osmotic flow and species transport phenomena in micro-mixer as well as convective heat/mass transfers for an array of electronic cooling devices. Parallel computations for global iterative solvers through Newton-Krylov methods devising AMR techniques have been carried out by PETSc solver interfacing with libMesh software package. Second, I will address my role of DOE funded SciDAC project currently participated. I have developed and implemented 3-D Discontinuous Galerkin Method (DGM) for large-scale fluid motions. I have integrated this 3-D dynamical core into the high-order method modeling environment (HOMME), which has currently proven to efficiently scale up to O(1000) parallel processors of IBM Blue Gene/L and IBM POWER5 p575 supercomputers. Parallelism employs a hybrid MPI/OpenMP design where its domain decomposition exploits the space-filling curve approach.