Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Vortex Particle Simulations of Aircraft Wake Instabilities on
Massively Parallel Architectures
with : M. Bergdorf, D. Rossinelli, P. Koumoutsakos A. Curioni, W. Andreoni (IBM Research Zurich)
CSE LabComputational Science & Engineering Laboratoryhttp://www.icos.ethz.ch/cse
Philippe Chatelain
VecPar’08
http://www.cse-lab.ethz.ch/ June 26 2008
Motivation
http://www.cse-lab.ethz.ch/ June 26 2008
Motivation
• COMPUTATION• Massively Parallel Simulations Using Particles
• IBM/BG: Distributed memory (104 - 105 CPUS), with relatively low RAM per node• Achieve sustained O(Tflops) performance
http://www.cse-lab.ethz.ch/ June 26 2008
Motivation
• COMPUTATION• Massively Parallel Simulations Using Particles
• IBM/BG: Distributed memory (104 - 105 CPUS), with relatively low RAM per node• Achieve sustained O(Tflops) performance
• PHYSICS• Optimization of aircraft wake decay
• aircraft & lift devices design
• High Reynolds DNS• Capture accurate decay of vortical flows• Physical insight - Basis for turbulent models
http://www.cse-lab.ethz.ch/ June 26 2008
• Lifting wing (with finite span)
• sheds a vortex sheet
• rolling up into a/several counter-rotating vortex pair(s)
• Vortex pair induces
• downwash
• rolling moment
➡ Safety hazard
➡ Currently, the limiting constraint on the scheduling of take-offs and landings in airports
Motivation: Aircraft Wakes
http://www.cse-lab.ethz.ch/ June 26 2008
Motivation: Aircraft Wakes
• Decay accelerationOutboard loading
Inboard loading
http://www.cse-lab.ethz.ch/ June 26 2008
Motivation: Aircraft Wakes
• Decay acceleration
• Modify lift distribution
Outboard loading
Inboard loading
http://www.cse-lab.ethz.ch/ June 26 2008
Motivation: Aircraft Wakes
• Decay acceleration
• Modify lift distribution
• change number and configuration of vortex pairs
Outboard loading
Inboard loading
http://www.cse-lab.ethz.ch/ June 26 2008
Motivation: Aircraft Wakes
• Decay acceleration
• Modify lift distribution
• change number and configuration of vortex pairs
• affects trajectories, growth rate of instabilities and decay
Outboard loading
Inboard loading
http://www.cse-lab.ethz.ch/ June 26 2008
Motivation: Aircraft Wakes
• Decay acceleration
• Modify lift distribution
• change number and configuration of vortex pairs
• affects trajectories, growth rate of instabilities and decay
• Actively perturb lift distribution, flow
• trigger instabilities
Outboard loading
Inboard loading
Leweke & Williamson, 1998
http://www.cse-lab.ethz.ch/ June 26 2008
Motivation: Aircraft Wakes
• Decay acceleration
• Modify lift distribution
• change number and configuration of vortex pairs
• affects trajectories, growth rate of instabilities and decay
• Actively perturb lift distribution, flow
• trigger instabilities
• Decay prediction
Outboard loading
Inboard loading
Leweke & Williamson, 1998
http://www.cse-lab.ethz.ch/ June 26 2008
Motivation: Aircraft Wakes
• Decay acceleration
• Modify lift distribution
• change number and configuration of vortex pairs
• affects trajectories, growth rate of instabilities and decay
• Actively perturb lift distribution, flow
• trigger instabilities
• Decay prediction
• Vortex tracking
Outboard loading
Inboard loading
Leweke & Williamson, 1998
http://www.cse-lab.ethz.ch/ June 26 2008
Outline
• Motivation
• Vortex Particle Method
• Implementation
• Medium Wavelength Instability
• Conclusions and Outlook
http://www.cse-lab.ethz.ch/ June 26 2008
Vortex Particle Method
http://www.cse-lab.ethz.ch/ June 26 2008
Vortex Particle Method• GOVERNING EQUATIONS
• Navier-Stokes, incompressible! · u = 0DuDt
= "1!!p + "!u
http://www.cse-lab.ethz.ch/ June 26 2008
Vortex Particle Method• GOVERNING EQUATIONS
• Navier-Stokes, incompressible
• Vorticity form
! · u = 0DuDt
= "1!!p + "!u
D!
Dt= (! ·!)u + "!!
! · u = 0! = !" u
http://www.cse-lab.ethz.ch/ June 26 2008
Vortex Particle Method• GOVERNING EQUATIONS
• Navier-Stokes, incompressible
• Vorticity form
• with velocity field
! · u = 0DuDt
= "1!!p + "!u
D!
Dt= (! ·!)u + "!!
! · u = 0
u = !" !!! = !"
" · ! = 0
! = !" u
http://www.cse-lab.ethz.ch/ June 26 2008
Vortex Particle Method• GOVERNING EQUATIONS
• Navier-Stokes, incompressible
• Vorticity form
• with velocity field
• DISCRETIZATION
! · u = 0DuDt
= "1!!p + "!u
D!
Dt= (! ·!)u + "!!
! · u = 0
u = !" !!! = !"
" · ! = 0
! = !" u
http://www.cse-lab.ethz.ch/ June 26 2008
Vortex Particle Method• GOVERNING EQUATIONS
• Navier-Stokes, incompressible
• Vorticity form
• with velocity field
• DISCRETIZATION
• Particles: position xp and strength
! · u = 0DuDt
= "1!!p + "!u
D!
Dt= (! ·!)u + "!!
! · u = 0
u = !" !!! = !"
" · ! = 0
!p =!
Vp
" dV ! "p Vp
! = !" u
http://www.cse-lab.ethz.ch/ June 26 2008
Vortex Particle Method• GOVERNING EQUATIONS
• Navier-Stokes, incompressible
• Vorticity form
• with velocity field
• DISCRETIZATION
• Particles: position xp and strength
• EVOLUTION EQUATIONS
! · u = 0DuDt
= "1!!p + "!u
D!
Dt= (! ·!)u + "!!
! · u = 0
u = !" !!! = !"
" · ! = 0
!p =!
Vp
" dV ! "p Vp
dxp
dt= u(xp)
d!p
dt=
!(" ·!)u(xp) + !!2"(xp)
"Vp
! = !" u
http://www.cse-lab.ethz.ch/ June 26 2008
Vortex Particle Method
http://www.cse-lab.ethz.ch/ June 26 2008
• Particle methods ARE NOT MESH-FREE
Vortex Particle Method
http://www.cse-lab.ethz.ch/ June 26 2008
• Particle methods ARE NOT MESH-FREE
➡ Particle distortion = loss of convergence
Vortex Particle MethodACCURATE
http://www.cse-lab.ethz.ch/ June 26 2008
• Particle methods ARE NOT MESH-FREE
➡ Particle distortion = loss of convergence
Vortex Particle MethodACCURATE
Particle Method Particle MethodDistortion handled by periodic regularization
Incompressible Euler Equations in 2D:(u-ω) formulation,evolution of an axis-symmetric vortex patchexact solution is steady
dxp
dt= u(xp, t)
d!p
dt= 0
http://www.cse-lab.ethz.ch/ June 26 2008
Remeshed Particle Methods
Qnewp =
!
p!
Qp!M(j h! xp!)
M(x)Interpolation Kernel• Moment conserving• Tensorial Product of 1D kernelse.g. high order B-splines
http://www.cse-lab.ethz.ch/ June 26 2008
• Remesh : reinitialize particles onto regular locations
Remeshed Particle Methods
Qnewp =
!
p!
Qp!M(j h! xp!)
M(x)Interpolation Kernel• Moment conserving• Tensorial Product of 1D kernelse.g. high order B-splines
http://www.cse-lab.ethz.ch/ June 26 2008
• Remesh : reinitialize particles onto regular locations
Remeshed Particle Methods
Qnewp =
!
p!
Qp!M(j h! xp!)
M(x)Interpolation Kernel• Moment conserving• Tensorial Product of 1D kernelse.g. high order B-splines
http://www.cse-lab.ethz.ch/ June 26 2008
• Remesh : reinitialize particles onto regular locations
Remeshed Particle Methods
Qnewp =
!
p!
Qp!M(j h! xp!)
M(x)Interpolation Kernel• Moment conserving• Tensorial Product of 1D kernelse.g. high order B-splines
http://www.cse-lab.ethz.ch/ June 26 2008
• Remesh : reinitialize particles onto regular locations
Remeshed Particle Methods
Qnewp =
!
p!
Qp!M(j h! xp!)
M(x)Interpolation Kernel• Moment conserving• Tensorial Product of 1D kernelse.g. high order B-splines
• Mesh also used for the efficient computation of Right-Hand Side
http://www.cse-lab.ethz.ch/ June 26 2008
Vortex Particle Mesh Methods
http://www.cse-lab.ethz.ch/ June 26 2008
• Hybrid scheme
• Mesh: RHS evaluations
Vortex Particle Mesh Methods
http://www.cse-lab.ethz.ch/ June 26 2008
• Hybrid scheme
• Mesh: RHS evaluations• Differential operators (F.D.)
• Fast Poisson solver (Fourier)
Vortex Particle Mesh Methods
!ij
uij
!!ij
http://www.cse-lab.ethz.ch/ June 26 2008
• Hybrid scheme
• Mesh: RHS evaluations• Differential operators (F.D.)
• Fast Poisson solver (Fourier)
• Particles only handle advection
Vortex Particle Mesh Methods
D!
Dt=
!!
!t+! · (!u)
!ij
uij
!!ij
http://www.cse-lab.ethz.ch/ June 26 2008
• Hybrid scheme
• Mesh: RHS evaluations• Differential operators (F.D.)
• Fast Poisson solver (Fourier)
• Particles only handle advection
• Particles and Mesh communicate through interpolation
Vortex Particle Mesh Methods
D!
Dt=
!!
!t+! · (!u)
!ij
uij
!!ij
!ij!p
http://www.cse-lab.ethz.ch/ June 26 2008
• Hybrid scheme
• Mesh: RHS evaluations• Differential operators (F.D.)
• Fast Poisson solver (Fourier)
• Particles only handle advection
• Particles and Mesh communicate through interpolation
Vortex Particle Mesh Methods
D!
Dt=
!!
!t+! · (!u)
!ij
uij
!!ij
!ij!p
uij(! ·!)uij
!!ij !!p
(! ·!)up
up
http://www.cse-lab.ethz.ch/ June 26 2008
• Hybrid scheme
• Mesh: RHS evaluations• Differential operators (F.D.)
• Fast Poisson solver (Fourier)
• Particles only handle advection
• Particles and Mesh communicate through interpolation
Vortex Particle Mesh Methods
D!
Dt=
!!
!t+! · (!u)
!ij
uij
!!ij
!ij!p
uij(! ·!)uij
!!ij !!p
(! ·!)up
up
•Stability properties of particle method unaffected: relaxed advection CFL
•Particle-Mesh interpolation: Efficient vectorizing loops, data alignment
http://www.cse-lab.ethz.ch/ June 26 2008
Implementation www.cse-lab.ethz.ch/software.html
http://www.cse-lab.ethz.ch/ June 26 2008
Implementation• Code is client of the PPM library,
Parallel Particle Mesh Library (Fortran 90, MPI)
www.cse-lab.ethz.ch/software.html
PPMers: I. Sbalzarini, J. Walther, M.Bergdorf, P. Chatelain, S. Hieber, E. Kotsalis, P. Koumoutsakos, F. Milde, M. Quack, B.
Hejazi Alhosseini
Sbalzarini et al, JCP 2006
http://www.cse-lab.ethz.ch/ June 26 2008
Implementation• Code is client of the PPM library,
Parallel Particle Mesh Library (Fortran 90, MPI)
• Here:
• Mesh-based domain decomposition
www.cse-lab.ethz.ch/software.html
PPMers: I. Sbalzarini, J. Walther, M.Bergdorf, P. Chatelain, S. Hieber, E. Kotsalis, P. Koumoutsakos, F. Milde, M. Quack, B.
Hejazi Alhosseini
Sbalzarini et al, JCP 2006
http://www.cse-lab.ethz.ch/ June 26 2008
Implementation• Code is client of the PPM library,
Parallel Particle Mesh Library (Fortran 90, MPI)
• Here:
• Mesh-based domain decomposition
• Data mappings
• particles: advected between subscommunication is flow dependent
• mesh values: ghosts for FD stencils
www.cse-lab.ethz.ch/software.html
PPMers: I. Sbalzarini, J. Walther, M.Bergdorf, P. Chatelain, S. Hieber, E. Kotsalis, P. Koumoutsakos, F. Milde, M. Quack, B.
Hejazi Alhosseini
Sbalzarini et al, JCP 2006
http://www.cse-lab.ethz.ch/ June 26 2008
Implementation• Code is client of the PPM library,
Parallel Particle Mesh Library (Fortran 90, MPI)
• Here:
• Mesh-based domain decomposition
• Data mappings
• particles: advected between subscommunication is flow dependent
• mesh values: ghosts for FD stencils
www.cse-lab.ethz.ch/software.html
PPMers: I. Sbalzarini, J. Walther, M.Bergdorf, P. Chatelain, S. Hieber, E. Kotsalis, P. Koumoutsakos, F. Milde, M. Quack, B.
Hejazi Alhosseini
Sbalzarini et al, JCP 2006
http://www.cse-lab.ethz.ch/ June 26 2008
Implementation• Code is client of the PPM library,
Parallel Particle Mesh Library (Fortran 90, MPI)
• Here:
• Mesh-based domain decomposition
• Data mappings
• particles: advected between subscommunication is flow dependent
• mesh values: ghosts for FD stencils
• Poisson solver
• 3D FFT
• Data Transposition uses PPM decompositions and mappings
www.cse-lab.ethz.ch/software.html
PPMers: I. Sbalzarini, J. Walther, M.Bergdorf, P. Chatelain, S. Hieber, E. Kotsalis, P. Koumoutsakos, F. Milde, M. Quack, B.
Hejazi Alhosseini
Sbalzarini et al, JCP 2006
http://www.cse-lab.ethz.ch/ June 26 2008
Implementation• Code is client of the PPM library,
Parallel Particle Mesh Library (Fortran 90, MPI)
• Here:
• Mesh-based domain decomposition
• Data mappings
• particles: advected between subscommunication is flow dependent
• mesh values: ghosts for FD stencils
• Poisson solver
• 3D FFT
• Data Transposition uses PPM decompositions and mappings
www.cse-lab.ethz.ch/software.html
PPMers: I. Sbalzarini, J. Walther, M.Bergdorf, P. Chatelain, S. Hieber, E. Kotsalis, P. Koumoutsakos, F. Milde, M. Quack, B.
Hejazi Alhosseini
Sbalzarini et al, JCP 2006
local:send
global:alltoallv
Chatelain et al, CMAME 2008
http://www.cse-lab.ethz.ch/ June 26 2008
Implementation• Client and library tuned for BG/L• Weak and strong scalabilities
• Drop at 4k: Poisson solver not sole culprit• Overhead in data mappings, latency in particle advection?
IBM T. J. Watson Center, Yorktown Heights, NJIBM Zurich Research Laboratory
Per-CPU problem sizeWithout Poisson solver
512 1024 2048 4096 8192 163840
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
NCPUS
ηwe
ak
2048 4096 8192 163840
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
NCPUS
ηst
rong
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mpe
r CPU
/ 10
6
Weak scal. Strong scal.
~ 400000 particles per CPU
http://www.cse-lab.ethz.ch/ June 26 2008
Implementation• Client and library tuned for BG/L• Weak and strong scalabilities
• Drop at 4k: Poisson solver not sole culprit• Overhead in data mappings, latency in particle advection?
IBM T. J. Watson Center, Yorktown Heights, NJIBM Zurich Research Laboratory
Per-CPU problem sizeWithout Poisson solver
512 1024 2048 4096 8192 163840
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
NCPUS
ηwe
ak
2048 4096 8192 163840
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
NCPUS
ηst
rong
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mpe
r CPU
/ 10
6
Weak scal. Strong scal.
~ 400000 particles per CPU
1010
http://www.cse-lab.ethz.ch/ June 26 2008
Implementation• Client and library tuned for BG/L• Weak and strong scalabilities
• Drop at 4k: Poisson solver not sole culprit• Overhead in data mappings, latency in particle advection?
IBM T. J. Watson Center, Yorktown Heights, NJIBM Zurich Research Laboratory
Per-CPU problem sizeWithout Poisson solver
512 1024 2048 4096 8192 163840
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
NCPUS
ηwe
ak
2048 4096 8192 163840
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
NCPUS
ηst
rong
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mpe
r CPU
/ 10
6
Weak scal. Strong scal.
~ 400000 particles per CPU
1010First prod. run
http://www.cse-lab.ethz.ch/ June 26 2008
• An interesting candidate...
• Counter-rotating vortex pairs
• generated by outboard loading
Aircraft Wakes
http://www.cse-lab.ethz.ch/ June 26 2008
• An interesting candidate...
• Counter-rotating vortex pairs
• generated by outboard loading
• or large horizontal tail plane
Aircraft Wakes
http://www.cse-lab.ethz.ch/ June 26 2008
• An interesting candidate...
• Counter-rotating vortex pairs
• generated by outboard loading
• or large horizontal tail plane
• Medium wavelength instability λ ~ span
Aircraft Wakes
http://www.cse-lab.ethz.ch/ June 26 2008
• An interesting candidate...
• Counter-rotating vortex pairs
• generated by outboard loading
• or large horizontal tail plane
• Medium wavelength instability λ ~ span
• secondary vortices wrap around primary ones
• reconnections of vortices with different circulations
Aircraft Wakes
Ortega, J.M. et al., JFM 2003
Durston D.A. et al., J Aircraft 2005
http://www.cse-lab.ethz.ch/ June 26 2008
• Long domain: length = long wavelength • Problem is periodic in all directions• Initial condition
• Vortices straight• Initiation by ambient noise
Medium Wavelength Instability
http://www.cse-lab.ethz.ch/ June 26 2008
• Long domain: length = long wavelength • Problem is periodic in all directions• Initial condition
• Vortices straight• Initiation by ambient noise
Medium Wavelength Instability
Physical parameters
• ReΓ = 6,000• Circulation ratio Γ2/Γ1 = -0.35• Span ratio b2/b1 = 0.5• Domain Lx = 10 b1
• Time = 0.35• Ambient white noise uRMS= 0.5%
CPU parameters
• IBM BlueGene/L• 4096 CPUS• 100 hours
Numerical parameters
• ~1.6 109 particles• 1024 x 768 x 2,048 grid• 10,000 time steps• RK3 Low-storage• 4th order FD
http://www.cse-lab.ethz.ch/ June 26 2008
Medium Wavelength Instability
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35−6
−5
−4
−3
−2
−1
0
t/to
E(k x)/E
0
kx = 0kx = 1kx = 8kx = 9kx =10kx =11kx =12
• Cross-flow energy
0 0.05 0.1 0.15 0.2 0.25 0.3 0.350.4
0.5
0.6
0.7
0.8
0.9
1
t/t0
E transverse(t)/E
transverse(0
)
Long
Medium
http://www.cse-lab.ethz.ch/ June 26 2008
• Energy spectra
• Mode λ/b1 = 0.86 - 0.943
• Experimental λ/b1 = 0.9 - 1.3• Bursts when Ω-loop feet come together
Medium Wavelength Instability
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35−6
−5
−4
−3
−2
−1
0
t/to
E(k x)/E
0
kx = 0kx = 1kx = 8kx = 9kx =10kx =11kx =12
• Cross-flow energy
0 0.05 0.1 0.15 0.2 0.25 0.3 0.350.4
0.5
0.6
0.7
0.8
0.9
1
t/t0
E transverse(t)/E
transverse(0
)
Long
Medium
http://www.cse-lab.ethz.ch/ June 26 2008
• Energy spectra
• Mode λ/b1 = 0.86 - 0.943
• Experimental λ/b1 = 0.9 - 1.3• Bursts when Ω-loop feet come together
• Visualization: volume rendering of |ω|• Parallel CPU-GPU ray-casting
Medium Wavelength Instability
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35−6
−5
−4
−3
−2
−1
0
t/to
E(k x)/E
0
kx = 0kx = 1kx = 8kx = 9kx =10kx =11kx =12
• Cross-flow energy
0 0.05 0.1 0.15 0.2 0.25 0.3 0.350.4
0.5
0.6
0.7
0.8
0.9
1
t/t0
E transverse(t)/E
transverse(0
)
Rossinelli D. et al., SIGGRAPH08
Long
Medium
http://www.cse-lab.ethz.ch/ June 26 2008
Instability : Linear phase t=0.21
http://www.cse-lab.ethz.ch/ June 26 2008
Instability : Linear phase t=0.21
http://www.cse-lab.ethz.ch/ June 26 2008
Instability : Reconnections t=0.25
http://www.cse-lab.ethz.ch/ June 26 2008
Instability : Reconnections t=0.25
http://www.cse-lab.ethz.ch/ June 26 2008
Instability: Propagation t=0.27
http://www.cse-lab.ethz.ch/ June 26 2008
Instability: Propagation t=0.27
http://www.cse-lab.ethz.ch/ June 26 2008
Instability: Decay t=0.35
http://www.cse-lab.ethz.ch/ June 26 2008
Instability: Decay t=0.35
Large Vortex Pair
Series of Vortex Rings
http://www.cse-lab.ethz.ch/ June 26 2008
Conclusions & Outlook
• Vortex Particle - Mesh Methods• Efficiency• Scalability on Massively Parallel Architectures...
• is good but could be better...• Mapping latency for particles• 3D FFT...
• Reference large-scale DNS of wake instabilities
http://www.cse-lab.ethz.ch/ June 26 2008
• Development
Conclusions & Outlook
http://www.cse-lab.ethz.ch/ June 26 2008
• Development
• Optimization/clean-up of global data mappings
Conclusions & Outlook
http://www.cse-lab.ethz.ch/ June 26 2008
• Development
• Optimization/clean-up of global data mappings
• Unbounded Poisson solvers in Fourier space
Conclusions & Outlook
http://www.cse-lab.ethz.ch/ June 26 2008
• Development
• Optimization/clean-up of global data mappings
• Unbounded Poisson solvers in Fourier space
• Applications
Conclusions & Outlook
http://www.cse-lab.ethz.ch/ June 26 2008
• Development
• Optimization/clean-up of global data mappings
• Unbounded Poisson solvers in Fourier space
• Applications
• Spatially developing flow
Conclusions & Outlook
http://www.cse-lab.ethz.ch/ June 26 2008
• Development
• Optimization/clean-up of global data mappings
• Unbounded Poisson solvers in Fourier space
• Applications
• Spatially developing flow
• forcing term accounts for wing lift, wake and jets
Conclusions & Outlook
http://www.cse-lab.ethz.ch/ June 26 2008
• Development
• Optimization/clean-up of global data mappings
• Unbounded Poisson solvers in Fourier space
• Applications
• Spatially developing flow
• forcing term accounts for wing lift, wake and jets
• capture roll-up, merger of vortex pairs
Conclusions & Outlook
http://www.cse-lab.ethz.ch/ June 26 2008
• Development
• Optimization/clean-up of global data mappings
• Unbounded Poisson solvers in Fourier space
• Applications
• Spatially developing flow
• forcing term accounts for wing lift, wake and jets
• capture roll-up, merger of vortex pairs
• Large Eddy Simulation
Conclusions & Outlook
http://www.cse-lab.ethz.ch/ June 26 2008
• Development
• Optimization/clean-up of global data mappings
• Unbounded Poisson solvers in Fourier space
• Applications
• Spatially developing flow
• forcing term accounts for wing lift, wake and jets
• capture roll-up, merger of vortex pairs
• Large Eddy Simulation
• Optimization of wake configurations with Evolution Strategies
Conclusions & Outlook