GPU-Accelerated Ab-Initio Simulations of Low-Pressure Turbines Richard D. Sandberg, Richard Pichler Aerodynamics and Flight Mechanics Research Group Vittorio Michelassi GE Global Research NVIDIA GPU Technology Conference, session on “Extreme-Scale Supercomputing with Titan Supercomputer”
26
Embed
GPU-Accelerated Ab-Initio Simulations of Low …on-demand.gputechconf.com/gtc/2014/presentations/S4756-gpu-sims-ab...GPU-Accelerated Ab-Initio ... basic questions regarding the physics
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
GPU-Accelerated Ab-Initio Simulations of Low-Pressure Turbines
Richard D. Sandberg, Richard Pichler Aerodynamics and Flight Mechanics Research Group
Vittorio Michelassi GE Global Research
NVIDIA GPU Technology Conference, session on “Extreme-Scale Supercomputing with Titan Supercomputer”
2 GPU-acceleration
3/26/2014
GE HPC Collaborations
Mill
ion
CP
U H
ou
rs
External collaborations, both in EU and USA, have
greatly accelerated HPC’s impact at GE ...
Year
3 GPU-acceleration
3/26/2014
HPT: Capture hot spot migration and
other temperature non-uniformity
Large-scale 3D unsteady flows
HPC impacting across components/disciplines...and driving compute cost.
Gas Turbines: Impacting Across Engine
Fan – Inlet:
Coupling and
predicting cross
wind behavior
Combustion: High fidelity for
Dynamics/Emissions
Jet noise: Move from test
screening to predict
LPT: Wake-transitional boundary layer
Interaction in unsteady environment
4 GPU-acceleration
3/26/2014
Motivation: why Low-Pressure-Turbine
Casing
Hub
Midspan Region
• Profile loss driven
Endwall Region
• Secondary loss drivenBlade 2
Vane 4
Blade 6
S
R
MID
SP
AN
Losses
EN
DW
ALL L
osses 1%
1% of LPT aerodynamic efficiency is worth 0.6-0.7% SFC
US airlines 35Bgallons in 2012
GE GT fleet alone $150B oil&gas/year
70+%
5 GPU-acceleration
3/26/2014
Aerodynamic design is assisted by tools based on modelling of turbulence However: modelling generally inaccurate
Want (model-free) DNS based on first principles, all length/time-scales must be resolved
Challenge: computational effort required increases approximately as ~Re3
Motivation
Spanwise vorticity for Ekman layer at Ret=1,241 Spanwise vorticity for Ekman layer at Ret=403
6 GPU-acceleration
3/26/2014
Motivation (why bother with DNS?)
• Reward for expending enormous computational resources:
wealth of reliable data free from modeling uncertainties used to answer
basic questions regarding the physics and modeling of a variety of flows
Advances in computing power mean DNS of Reynolds
numbers/ geometries of interest now possible
TOP500 Global Trends
from Jack Dongarra,
IESP
7 GPU-acceleration
3/26/2014
Aim of this presentation:
Introduce novel Navier-Stokes solver purposely developed to exploit
modern HPC architectures for DNS
1) Present key ingredients for an efficient algorithm
2) Performance study of original hybrid OMP/MPI code
3) Porting of code to GPU-accelerated architecture (Titan) using Open-ACC
4) Performance study of Open-ACC code
5) Preliminary results of LPT at real operating conditions