PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS:...
Transcript of PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS:...
![Page 1: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/1.jpg)
PALADINS:Scalable Time-Adaptive Algebraic Splitting
and Preconditioners for the Navier-Stokes Equations
U. Villa, A. VenezianiDepartment of Mathematics and Computer Science
EMORY UNIVERSITY
Atlanta, GA, US
2013 SIAM ConferenceOn Computational Science and Engineering
Boston, 27th February 2013
![Page 2: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/2.jpg)
• Introduction– Motivations– Basic settings– Incremental pressure schemes
• Algebraic Splittings– Inexact LU block factorizations– High Order Yosida Schemes
• Adaptivity– Local splitting error analysis– A posteriori error estimators– An application to blood flow problems
• Scalability– Strong scaling test– Weak scaling test
• Conclusions
Outline:
![Page 3: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/3.jpg)
Some applications of INS feature sequences of fast and slow transients
Motivations:
Time adaptivity can reduce CPU times in these applications… An effective a-posteriori error estimator is however mandatory
Standard a-posteriori error estimator requires:• Complex space-time error estimator or• The comparison of two numerical solutions obtained with different
accuracy time discretizations (eg. Adams–Bashforth/BDF2, as in Kay, Gresho, Griffiths, Silvester, 2008)
Blood flow dynamics:• fast transients during systole• slower dynamics during diastole
Algebraic splittings of velocity/pressure can provide effective estimator as a by product of the computations.
![Page 4: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/4.jpg)
Basic settings:
Incompressible UnsteadyNavier-Stokes Equations
Space: Galerkin methods LBB conditions fulfilled
Time: BDFq schemes (q≤3)
Discretization
At each time level t=tn+1 we need to solve the system:
being A the stiffness matrix (viscous stresses + convection terms)
![Page 5: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/5.jpg)
Incremental Pressure Schemes:
is the pressure extrapolated from previous time steps, and
At each time step we write , where
• For velocity/pressure splitting:• Incremental pressure schemes improve the accuracy in time.• High order extrapolation in time might reduce the stability.
• For Schur-Complement/Monolithic solutions:• Incremental pressure provides a good initial guess.• High order extrapolation in time does not affect stability.
A. Prohl – Projection and quasi-compressibility methods for solving the incompressible Navier-Stokes equations, Wiley Teubner Advances in Num. Math., 1997
M. Henriksen and J. Holmen – Algebraic Splitting for Incompressible Navier-Stokes Equations, J. Comp.
For example if s=1 then and if s = 2 then
The incremental pressure formulation reads
![Page 6: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/6.jpg)
• Introduction– Motivations– Basic settings– Incremental pressure schemes
• Algebraic Splittings– Inexact LU block factorizations– High Order Yosida Schemes
• Adaptivity– Local splitting error analysis– A posteriori error estimators– An application to blood flow problems
• Scalability– Strong scaling test– Weak scaling test
• Conclusions
Outline:
![Page 7: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/7.jpg)
Inexact LU block factorization:
F and G appropriate approximations of C 1, Q is such that DGDT DFDTQ is small
Neumann expansion:
Mass preserving scheme: F=G=H, Q=INp (Algebraic Chorin Temam, Perot ’93)
Momentum preserving sch.: G=C-1, F=H (Q=INp: Yosida, Quarteroni, Saleri, Veneziani,’99)
H
Higher order schemes build a sequence of Qq such that:
provided Δt is small enough.
being and
F. Saleri, A. Veneziani – Pressure correction algebraic splitting methods for the incompressible Navier-Stokes equations – SIAM J. Num. Anal. (2006)
![Page 8: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/8.jpg)
Inexact LU block factorization:
Velocity splitting errors Pressure splitting errors
Higher order schemes build a sequence of Qq such that:
being and
A. Veneziani, A Note on the Consistency and Stability Properties of Yosida Fractional Step Schemes for the Unsteady Stokes Equations, SIAM J. Numer. Anal.,2009
![Page 9: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/9.jpg)
High order Yosida schemes:
//Pressure corrections
z(i) = ZeroVector(dim_P);
//Temporanely data structures
zz(i,j) = ZeroVector(dim_U);
dzz(i,j)= ZeroVector(dim_P);
Solve: S z(0) = rhs;
for(i=0; i<q; ++i)
zz(i,0) = -H A H DT z(i);
dzz(i,0) = D zz(i,0);
cc = dzz(i,0);
for(j=1; j<1+i; ++j)
zz(i-j,j) = - H A zz(i-j,j-1);
dzz(i-j,j) = D zz(i-j,j);
cc += dzz(i-j,j);
Solve: S z(i+1) = cc;
P = sum(z);
Norms of the pressure corrections zi
A. Veneziani, U. Villa – ALADINS: an ALgebraic splitting time ADaptive solver for the Incompressible Navier-Stokes equations, J. Comput. Phys. (2013)
Algorithm to apply (SQq)-1 to a vector
timestep
Computational cost for each correction step:- Three mat-vec in the velocity space- One linear solve with the spd matrix S
![Page 10: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/10.jpg)
Eigenvalues of (SQp)-1
High Order Yosida as Preconditioner:
![Page 11: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/11.jpg)
Eigenvalues of (SQp)-1
High Order Yosida as Preconditioner:
![Page 12: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/12.jpg)
High Order Yosida as Preconditioner:
Eigenvalues of (SQp)-1
![Page 13: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/13.jpg)
• Introduction– Motivations– Basic settings– Incremental pressure schemes
• Algebraic Splittings– Inexact LU block factorizations– High Order Yosida Schemes
• Adaptivity– Local splitting error analysis– A posteriori error estimators– An application to blood flow problems
• Scalability– Strong scaling test– Weak scaling test
• Conclusions
Outline:
![Page 14: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/14.jpg)
Local Splitting error analysis:
Unsplit formulation
Split formulation
-
=
Local splitting error
Stokes System, incremental pressure approach*
*The non incremental approach has been analyzed in P. Gervasio. SIAM J. Numer. Anal., 2008
and
A. Veneziani, U. Villa – ALADINS: an ALgebraic splitting time ADaptive solver for the Incompressible Navier-Stokes equations, J. Comput. Phys. (2013)
![Page 15: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/15.jpg)
Local Splitting error analysis:
Non incremental method
Womersley analytical solution
Vel
oci
ty s
plit
tin
g er
ror
Pre
ssu
re s
plit
tin
g er
ror
![Page 16: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/16.jpg)
Local Splitting error analysis:
Incremental method (s=1)
Womersley analytical solution
Vel
oci
ty s
plit
tin
g er
ror
Pre
ssu
re s
plit
tin
g er
ror
![Page 17: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/17.jpg)
1. Yosida(q) – Yosida(q-1):
• Splitting based adaptivity (conditionally stable)
• The last pressure increment zq provides the error estimator.
2. Monolithic-Yosida(q-1) :
• Preconditioning based adaptivity (unconditionally stable)
• The difference between the split and unsplit solution provides the error estimator
A posteriori error estimators:
![Page 18: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/18.jpg)
1. Yosida(q) – Yosida(q-1):
• Splitting based adaptivity (conditionally stable)
• The last pressure increment zq provides the error estimator.
2. Monolithic-Yosida(q-1) :
• Preconditioning based adaptivity (unconditionally stable)
• The difference between the split and unsplit solution provides the error estimator
then we pick where and
A posteriori error estimators:
h.o.t.
Assume we require an accuracy τ for the absolute pressure error
A. If χ < 1 reject the time stepB. If χ ≥ 1 accept the time step
h.o.t.
![Page 19: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/19.jpg)
At each time step we solve the coupled system in the velocity u and the pressure increment δp.
Monolithic – Yosida(q-1) Adaptivity:
As a left preconditioner we use the lower triangular part of the Yosida(q-1) splitting
Let the first preconditioned residual.
a posteriori error estimator:
Note: the High Order Yosida Preconditioner SQ1 is equivalent to the Least Square Commutator preconditioner by Elman (SIAM J. Sci. Comput., 1999)
![Page 20: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/20.jpg)
Blood flow application:
Adaptivity:Monolithic – Yosida 1 error estimatorSecond order error estimator
Discretization:TIME: BDF2 with incremental pressure (s=1)SPACE: Inf-sup compatible P1Bubble-P1 FE
Reynolds Womersley
300 21
Real geometry, physiological conditions
Pressure peak Flux peak
Inflow pressure and flux
![Page 21: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/21.jpg)
Blood flow application:
![Page 22: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/22.jpg)
Blood flow application:
Pressure norm; Time Step Velocity norm; Time Step
Wall shear stress; Time Step
Steps perheart beat
Non adaptive 834
Adaptive 221
Speed-up 3.75
![Page 23: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/23.jpg)
• Introduction– Motivations– Basic settings– Incremental pressure schemes
• Algebraic Splittings– Inexact LU block factorizations– High Order Yosida Schemes
• Adaptivity– Local splitting error analysis– A posteriori error estimators– An application to blood flow problems
• Scalability– Strong scaling test– Weak scaling test
• Conclusions
Outline:
![Page 24: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/24.jpg)
At each time step we solve the coupled system in the velocity u and the pressure p
with preconditioned GMRES iterations (Belos)
The block upper-triangular variant of the High Order Yosida Preconditioner
is applied inexactly using the AMG preconditioners available in ML(Smoothed Aggregation and Symmetric Gauss-Seidel smoothers) for C and S
Energy minimization prolongation and unsmoothed aggregation are used to cope with the non-symmetry of C
Solution of the saddle-point system:
![Page 25: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/25.jpg)
Strong Scalability Test:
Simulation of blood flow in a giant aneurysm on the internal carotid artery.
Benchmark proposed in the CFD Challenge Workshop at ASME 2012.
T. Passerini, J. Slawinski, U. V., A. Veneziani, V. Sunderam – Experiences with a computational fluid dynamics code on clouds, grids, and on-premise resources. (submitted to JPDC 2012)
Space discretization: P1Bubble-P1 elements (≈ 3M unknowns).Time discretization: BDF2 (timestep 0.01s).
![Page 26: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/26.jpg)
Weak Scalability Test:
Ethier-Steinman Benchmark (1994)Unstructured Tetrahedral Mesh (Netgen)Low Reynolds number (approx 100)
Space discretization
Taylor Hood P2-P1 FE Mini Element P1B-P1 FE
Second order approx of velocity First order approx of velocity
Denser FE matrices Sparser matrices
No mass lumping Accurate mass lumping
Convective term treatment
Semi-implicit Explicit
Non symmetric momentum matrix Symmetric momentum matrix
Add grad-div stabilization Block diagonal momentum matrix
Time-step proportional to the mesh diameter (accuracy and stability)
![Page 27: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/27.jpg)
Weak Scalability Test:
P2-P1 Finite Elements (Consistent Velocity Mass Matrix)
Stopping criterion: relative norm of the residual less than 10-9
![Page 28: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/28.jpg)
Weak Scalability Test:
P2-P1 Finite Elements (Consistent Velocity Mass Matrix)
Stopping criterion: relative norm of the residual less than 10-9
![Page 29: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/29.jpg)
Weak Scalability Test:
P1B-P1 Finite Elements (Lumped Velocity Mass Matrix)
Stopping criterion: relative norm of the residual less than 10-9
![Page 30: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/30.jpg)
Weak Scalability Test:
P1B-P1 Finite Elements (Lumped Velocity Mass Matrix)
Stopping criterion: relative norm of the residual less than 10-9
![Page 31: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/31.jpg)
Weak Scalability Speedup:
![Page 32: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/32.jpg)
• Introduction– Motivations– Basic settings– Incremental pressure schemes
• Algebraic Splittings– Inexact LU block factorizations– High Order Yosida Schemes
• Adaptivity– Local splitting error analysis– A posteriori error estimators– An application to blood flow problems
• Scalability– Strong scaling test– Weak scaling test
• Conclusions
Outline:
![Page 33: PALADINS - ta.twi.tudelft.nlta.twi.tudelft.nl/nw/users/vuik/siam_cse_2013/UVILLA.pdf · PALADINS: Scalable Time-Adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes](https://reader036.fdocuments.us/reader036/viewer/2022071217/604d315569a82b4a22582000/html5/thumbnails/33.jpg)
• Incremental pressure methods improve the accuracy of the splitting.
• High order Yosida splittings provide an effective time adaptivity error
estimator as a by-product of the computation.
• Schur complement/Monolithic adaptive schemes allows selection of larger
time-step due to their unconditionally stability.
• High order Yosida splittings are optimal preconditioners for the unsteady NSE.
• (P)ALADINS is a (Parallel) ALgebraic ADaptive Incompressible Navier-Stokes
Solver, based on algebraic splitting of velocity and pressure.
• Good strong and weak scaling properties in parallel when the local problem
size is large enough using Trilinos (ML, Belos).
Conclusions: