Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.arcetri.astro.it/igm50/posters_all/schneider.pdf Дата изменения: Fri Jul 24 10:51:42 2015 Дата индексирования: Sun Apr 10 03:14:09 2016 Кодировка: Поисковые слова: annular solar eclipse

Cholla: A New Code for Astrophysical Hydrodynamics
Evan Schneider and Brant Robertson
University of Arizona

Introduction
CHOLLA stands for Computational Hydrodynamics On paraLLel Architectures. Despite increasingly powerful CPUs, many astrophysical simulations remain bound by computational expense. Problems requiring static mesh hydrodynamics calculations often fall within this category. To improve upon this situation, we have developed Cholla, a code that takes full advantage of newer, extremely parallel GPU-based computer architectures. Using GPUs allows us to perform high resolution simulations quickly (and for a fraction of the cost of CPU-based methods). We are currently using Cholla to probe questions about galactic outflows as well as ISM turbulence. We plan to make the code public soon.
·

Code Specifics
Cholla is a new three-dimensional hydrodynamics code that harnesses the power of GPUs to accelerate astrophysical simulations. Details: static mesh Eulerian code, unsplit Corner Transport Upwind integrator (Gardiner & Stone, 2008) or Van Leer predictor-corrector scheme (Stone & Gardiner, 2009), exact and Roe Riemann solvers, up to third-order spatial reconstruction with several slope limiting options, and optically thin radiative cooling. Cholla is fast (5 - 20 million cell updates per GPU second), MPI parallelized, and shows demonstrated scaling to > 16000 GPUs.

Cloud-Wind Interactions
What happens if clouds are not treated as constant density, uniform spheres?

·

·

GPU-based = Fast hydrodynamics
each thread does hydro calculations for one cell

Weak Scaling on Titan
Total runtime (excluding initialization) Boundary conditions CTU (all GPU functions) 100

threads {0...n}

Shared Memory

Block 0,0

Block 1,0

Block 2,0

Block 3,0

grid of thread blocks Block 0,1 Block 1,1 Block 2,1 Block 3,1

Time (s)

Titan is the flagship supercomputer at Oak Ridge National Laboratory, with 18,688 GPUequipped compute nodes. In preliminary tests, Cholla achieves nearly ideal weak scaling on tests with up to 81923 (550 billion) cells.

·

Most numerical studies of galactic outflows have modeled the cool gas as constant density, spherical clouds (e.g. Scannapieco et al., 2015). We test a more realistic model in which the cool clouds start with a lognormal density structure as a result of their turbulent nature. We demonstrate that turbulent clouds get disrupted and dissipate much more quickly than spheres with equivalent mass and mean density.
References: Schneider, Evan E. & Robertson, Brant E. (2015). Cholla: A New Massively-Parallel Hydrodynamics Code For Astrophysical Simulation. The Astrophysical Journal Supplement, 217(2), 24-58.

·

·
10

CPU

GPU Global Memory

1 1

10 100 1000 MPI processes (GPUs)

10000

This work funded in part by NSF grants No. DGE-1143953 & No. 1228509.