Cha-Cha Days 2010. September 24-26, 2010. Two-level Additive Schwarz Preconditioners for a Weakly Over-Penalized Symmetric Interior. Penalty Method.
Cha-Cha Days 2010 September 24-26, 2010
Two-level Additive Schwarz Preconditioners for a Weakly Over-Penalized Symmetric Interior Penalty Method
Eun-Hee Park Mathematics & Center for Computation and Technology
Joint work with A.T. Barker, S.C. Brenner, and L.-Y. Sung
Cha-Cha Days 2010
Eun-Hee Park – 1
Outline
Domain Decomposition Methods Weakly Over-Penalized Symmetric Interior Penalty (WOPSIP) Method Two-level Additive Schwarz Preconditioners Numerical Results
Cha-Cha Days 2010
Eun-Hee Park – 2
Outline
Domain Decomposition Methods Weakly Over-Penalized Symmetric Interior Penalty (WOPSIP) Method Two-level Additive Schwarz Preconditioners Numerical Results
Cha-Cha Days 2010
Eun-Hee Park – 2
Outline
Domain Decomposition Methods Weakly Over-Penalized Symmetric Interior Penalty (WOPSIP) Method Two-level Additive Schwarz Preconditioners Numerical Results
Cha-Cha Days 2010
Eun-Hee Park – 2
Outline
Domain Decomposition Methods Weakly Over-Penalized Symmetric Interior Penalty (WOPSIP) Method Two-level Additive Schwarz Preconditioners Numerical Results
Cha-Cha Days 2010
Eun-Hee Park – 2
Outline
Domain Decomposition Methods Weakly Over-Penalized Symmetric Interior Penalty (WOPSIP) Method Two-level Additive Schwarz Preconditioners Numerical Results
Cha-Cha Days 2010
Eun-Hee Park – 2
Domain Decomposition Methods Why DDM ? What’s DDM ? WOPSIP Method Two-level Additive Schwarz Method Numerical Results
Domain Decomposition Methods
Cha-Cha Days 2010
Eun-Hee Park – 3
Why DDM ?
Domain Decomposition Methods Why DDM ? What’s DDM ?
Toy Problem
WOPSIP Method Two-level Additive Schwarz Method Numerical Results
Cha-Cha Days 2010
−Δu
=
f
in Ω
u
=
0
on ∂Ω
Eun-Hee Park – 4
Why DDM ?
Domain Decomposition Methods Why DDM ? What’s DDM ?
Toy Problem
WOPSIP Method Two-level Additive Schwarz Method 1
Numerical Results
0.8 0.6 1 0.4
0.8 0.6
0.2
0.4 0
0.2 0
Solved in the blink of an eye !
Cha-Cha Days 2010
−Δu
=
f
in Ω
u
=
0
on ∂Ω
Eun-Hee Park – 5
Why DDM ?
Domain Decomposition Methods Why DDM ? What’s DDM ?
Real Application
WOPSIP Method
◆ Large scale problems ◆ How to solve in limited
Two-level Additive Schwarz Method Numerical Results
time and memory ?
Domain Decomposition ◆ Divide into many of
smaller problems ◆ Solve subdomain prob-
lems in parallel
- Solution for limitation of memory and computing time Cha-Cha Days 2010
Eun-Hee Park – 6
What’s DDM ?
Classical Schwarz Alternating Method (Schwarz, 1869) ◆ Iterative Method
Domain Decomposition Methods Why DDM ? What’s DDM ? WOPSIP Method
- Classical solutions smooth regions
on
Two-level Additive Schwarz Method
Ω
non-
Numerical Results
◆ Alternating Procedure Given an initial u0 , solve successively: −Δu2k+1 = f u2k+1 = u2k −Δu2k+2 = f u2k+2 = u2k+1 ◆ Numerical Method ?
Cha-Cha Days 2010
in Ω1
Ω1
on ∂Ω1 , in Ω2 on ∂Ω2 . Ω2
Eun-Hee Park – 7
What’s DDM ?
Domain Decomposition Methods Why DDM ? What’s DDM ?
Discretized Problem k+1
u
Ah u = f k −1 k f − Ah u =u + P
Ω
WOPSIP Method Two-level Additive Schwarz Method
P −1 Ah u = P −1 f
Numerical Results
DD-based Preconditioner
[Dryja & Widlund (1987)]
- Additive Schwarz Preconditioner
Ω1
Ω2
k+1
u
k
=u +
A−1 1
+
−1 A2
k
f − Ah u
- Two-level Additive Schwarz Preconditioner Cha-Cha Days 2010
Eun-Hee Park – 8
Domain Decomposition Methods WOPSIP Method Two-level Additive Schwarz Method Numerical Results
WOPSIP Method
Cha-Cha Days 2010
Eun-Hee Park – 9
Discontinuous Galerkin (DG) methods
Why DG methods ◆ Need for Nonconforming Meshes
- Adaptive methods e.g. L-shaped domain
Poisson Model Problem
Cha-Cha Days 2010
−Δu
=
f
in Ω
u
=
0
on ∂Ω
Eun-Hee Park – 10
Discontinuous Galerkin (DG) methods
DG methods For v, w ∈ Vh , a± h (v, w) = ±
T ∈Th
e∈Eh
T
∇v · ∇w dx −
e∈Eh
e
{{∇v}} · [ w]] ds
η {∇w}} · [ v]] ds + [ v]] · [ w]] ds, e{ |e| e e∈E h
where
Vh = v ∈ L2 (Ω) : v|T ∈ P1 (T ) ∀T ∈ Th .
◆ a− h : SIPG (Symmetric Interior Penalty Galerkin) method [Douglas & Dupont (1976), Wheeler (1978), Arnold (1982)]
◆ a+ h : NIPG (Nonsymmetric Interior Penalty Galerkin) method [Rivi` ere, Wheeler & Girault (1999)]
Cha-Cha Days 2010
Eun-Hee Park – 11
Comparison: SIPG & NIPG
Domain Decomposition Methods
SIPG ◆ Correct error estimates in both energy norm and L2 norm ◆ Stability: η > η0 for a sufficiently large η0
WOPSIP Method Two-level Additive Schwarz Method Numerical Results
NIPG ◆ Correct error estimate only in energy norm ◆ Stability: η > η0 for an arbitrary η0
WOPSIP ◆
Weakly Over-Penalized Symmetric Interior Penalty
◆
Brenner, Owens & Sung (2008)
Cha-Cha Days 2010
Eun-Hee Park – 12
WOPSIP method
WOPSIP Formulation To find uh ∈ Vh such that, for any vh ∈ Vh , 1 0 0 ∇uh · ∇vh dx + η Π [ u ] · Π h e e [ vh ] ds = (f, vh )L2 (Ω) , 3 |e| T e T ∈T e∈E h
h
where Π0e is the orthogonal projection from [L2 (e)]2 onto [P0 (e)]2 .
Advantages ◆ SPD ◆ No need for tuning of a penalty parameter (η = 1) ◆ Correct error estimates in both energy norm and L2 norm ◆ Meshes with hanging nodes ◆ Simple programming ◆ Intrinsic parallelism
Cha-Cha Days 2010
Eun-Hee Park – 13
Domain Decomposition Methods WOPSIP Method Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis
Two-level Additive Schwarz Method
Cha-Cha Days 2010
Numerical Results
Eun-Hee Park – 14
Two-level Additive Schwarz Preconditioners
Domain Decomposition Methods
Linear System
WOPSIP Method
Ah uh = f
Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis
Iterative Solver
Numerical Results
◆ Preconditioned Conjugate Gradient Method
B −1/2 Ah B −1/2 v = B −1/2 f,
◆ uh − uk Ah
Cha-Cha Days 2010
v = B 1/2 uh
k √ −1 κ(B Ah )−1 ≤2 √ uh − u0 Ah κ(B −1 Ah )+1
Eun-Hee Park – 15
Two-level Additive Schwarz Preconditioners (cont.)
Two-level Additive Schwarz Preconditioner
Domain Decomposition Methods WOPSIP Method
B
=
Bc + Bf
=
h h t IH AH −1 (IH )
+
N
Ii Ah,i −1 Iit
i=1
Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis Numerical Results
◆ AH : Coarse Solver (associated w/ VH ) ◆ Ah,i : Subdomain Solvers (associated w/ Vi ) ◆ Ii : Natural injection operator h ◆ IH : Intergrid transfer operator
Cha-Cha Days 2010
Eun-Hee Park – 16
Two-level Preconditioners (cont.)
Subdomain Solvers
Domain Decomposition Methods
Ah,i : Vi → Vi defined by ∇u · ∇v dx + Ah,i u, v =
WOPSIP Method
T ∈Thi
T
i e∈Eh
1 |e|3
e
Π0e [ u]] · Π0e [ v]] ds
Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis Numerical Results
Vi = {v ∈ Vh : v = 0 on Ω \ Ωi }
TH
Cha-Cha Days 2010
Th
˜ i → Ωi Ω
Eun-Hee Park – 17
Two-level Preconditioners (cont.)
Domain Decomposition Methods
Coarse Solver ◆ VH,1 :
AH v, w =
◆ VH,2 :
AH v, w =
◆ VH,3 :
AH v, w =
◆ VH,4 :
AH v, w = T ∈TH
conforming P1
Cha-Cha Days 2010
T
T ∈TH T ∈TH
WOPSIP Method
T T
1 e∈EH |e|
∇v · ∇w dx ∇v · ∇w dx 0 0 e Πe [ v]] · Πe [ w]] ds
Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis Numerical Results
∇v · ∇w dx +
1 e∈EH |e|
non-conforming P1 discontinuous P0
e
Π0e [ v]] · Π0e [ w]] ds
discontinuous P1
Eun-Hee Park – 18
Two-level Preconditioners (cont.)
Domain Decomposition Methods
Preconditioned System BAh =
h h t IH AH −1 (IH ) +
N
Ii Ah,i −1 Iit
i=1
WOPSIP Method
Ah
Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis Numerical Results
Intergrid Transfer Operator h IH : VH → Vh such that h h v, IH v ≤ CAH v, v Ah IH
Cha-Cha Days 2010
∀v ∈ VH .
Eun-Hee Park – 19
Intergrid Transfer Operator
Difference in Scales For a piecewise constant function vH ∈ VH,3 ⊂ VH,4 ⊂ Vh , Ah vH , vH
=
e∈Eh
AH vH , vH
=
E∈EH
1 |e|3 1 |E|
0 2 |Π [ v ] | ds H e e
0 2 |Π [ v ] | ds H E E
Domain Decomposition Methods WOPSIP Method Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis Numerical Results
Edge by Edge ◆ Fine Solver
1 |e|3
◆ Coarse Solver
1 |E| Cha-Cha Days 2010
e
|Π0e [ vH ] |2 ds ∼ O
E
1 |e|2
|Π0E [ vH ] |2 ds ∼ O(1) Eun-Hee Park – 20
Need for Enriching Operator BAh =
h h t IH AH −1 (IH ) +
t Ii A−1 h,i Ii
i=1
CAh =
N
I0 AH −1 (I0 )t +
N
Domain Decomposition Methods
Ah
t Ii A−1 h,i Ii
Ah
i=1
CAh ignoring difference in scales h 2−6 2−7 2−8 2−9
N =4 CG its. 184 303 491 803
κ(CAh ) 1.14·105 1.36·106 1.72·107 2.43·108
N = 32 CG its. 184 288 494 936
WOPSIP Method Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis Numerical Results
κ(CAh ) 1.11·105 1.38·106 1.92·107 2.93·108
- I0 : Interpolation without enrichment - Discontinuous P 1 coarse space - Coarse mesh size H = 2−5 Cha-Cha Days 2010
Eun-Hee Park – 21
Intergrid Transfer Operator (cont.)
Domain Decomposition Methods
Enriching Operator EH : VH,4 → VH,4 ∩ H01 (Ω)
WOPSIP Method Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis
◆ Averaging Operator
At a vertex p, we define 1 (EH v)(p) = v|T (p), |Tp | T ∈T
Numerical Results
p
where |Tp | is the number of triangles sharing p as a vertex. ◆ Ran(EH ) ⊂ C 0 (Ω)
For each e ∈ Eh , |Π0e [ EH vH ] |2 ds = 0. e
cf. Cha-Cha Days 2010
1 |e|3
0 e |Πe
[ vH
] |2
ds = O
1 |e|2
Eun-Hee Park – 22
Intergrid Transfer Operator (cont.)
Ih H : Coarse Space → Fine Space h IH = Πh ◦ EH
◆ EH : Enriching operator ◆ Πh : Crouzeix-Raviart interpolation operator - Πh : H 1 (Ω) → Vh (Πh v)|T = ΠT (v|T ) - ΠT : H 1 (T ) → P1 (T ) 1 (ΠT v)(me ) = |e|
Cha-Cha Days 2010
∀T ∈ Th
e
v ds
∀e ∈ ET
Eun-Hee Park – 23
Convergence Analysis
Requirement on Ih H : VH → Vh
Domain Decomposition Methods
h h Ah IH v, IH v
≤
CAH v, v
h v − IH v L2 (Ω)
≤
CHAH v, v
∀v ∈ VH ∀v ∈ VH
Cf. Abstract Schwarz theory [Dryja & Widlund (1987)]
WOPSIP Method Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis Numerical Results
Theorem The condition number of the preconditioned system BAh is estimated as
κ(BAh ) ≤ C 1 +
H δ
,
where C is a constant independent of H, h, δ and N . - H: coarse mesh size - δ: overlap width - N : number of subdomains [Barker, Brenner, Park & Sung, to appear]
Cha-Cha Days 2010
Eun-Hee Park – 24
Convergence Analysis
H κ(BAh ) ≤ C 1 + δ
Domain Decomposition Methods WOPSIP Method Two-level Additive Schwarz Method Two-level Preconditioner Intergrid Transfer Operator Convergence Analysis
- Linear growth - Optimal preconditioner if H/δ ∼ 1
Numerical Results
- Scalable algorithm - B: preconditioner with enriching process for WOPSIP method
Remark Two-level Schwarz preconditioners for super penalty DGM 1 κ(PT LAS Ah ) ≤ C 2 h
H 1+ h
[Antonietti & Ayuso (2009)]
Cha-Cha Days 2010
Eun-Hee Park – 25
Domain Decomposition Methods WOPSIP Method Two-level Additive Schwarz Method Numerical Results
Numerical Results
Cha-Cha Days 2010
Eun-Hee Park – 26
Numerical Results
Poisson Problem
Computing Environment ◆ LONI Supercomputers
Find a FE approximation uh of the exact solution u such that −Δu
=
f
in Ω
u
=
0
on ∂Ω
Domain Decomposition Methods WOPSIP Method Two-level Additive Schwarz Method Numerical Results
where Ω = (0, 1) × (0, 1) ⊂ R2 and u(x, y) = xy(1 − x)(1 − y). (Louisiana Optical Network Initiative)
Cha-Cha Days 2010
Eun-Hee Park – 27
One-level Preconditioner
Domain Decomposition Methods
BAh = (Bf + Bc ) Ah
WOPSIP Method
W/O Coarse Solver
Two-level Additive Schwarz Method
Bf Ah =
N
Numerical Results
Ii Ah,i −1 Iit
Ah
i=1
N 2 4 8 16 32
CG its. 42 65 73 90 109
κ(Bf Ah ) 223.39 739.55 876.95 1.14·103 1.47·103
Table 1: N( of subdomains), fine mesh h = 2−8 , overlap width δ=h Cha-Cha Days 2010
Eun-Hee Park – 28
Comparison: CG iteration counts
N 1 2 4 8 16 32 64 128 256 512
one-level 1 42 65 73 90 109 135 149 193 220
VH,1 2 11 11 11 11 11 11 11 11 10
VH,2 3 14 15 14 14 14 14 15 14 14
VH,3 4 13 14 14 14 14 15 15 15 15
VH,4 8 13 13 14 14 14 13 12 14 14
Domain Decomposition Methods WOPSIP Method Two-level Additive Schwarz Method Numerical Results
Table 2: H/δ = 8, fine mesh h = 2−8 , coarse mesh H = 2−5 , overlap: δ = h
Cha-Cha Days 2010
Eun-Hee Park – 29
VH,4 : Dependency on H and δ
Domain Decomposition Methods
Coarse mesh size H H 2−4 2−5 2−6 2−7
N =4 CG its. 19 13 10 8
wall clock time 3.9 3.6 4.8 12.7
N = 32 CG its. 20 14 10 7
WOPSIP Method
wall clock time .96 .91 1.0 1.8
Two-level Additive Schwarz Method Numerical Results
Table 3: fine mesh h = 2−8 , overlap width δ = h Overlap width δ δ 0 h 2h 3h
N =4 CG its. 22 12 9 10
wall clock time 4.7 3.5 3.2 3.4
N = 32 CG its. 17 12 8 9
wall clock time 1.3 1.0 .72 .70
Table 4: fine mesh h = 2−8 , coarse mesh H = 2−5 Cha-Cha Days 2010
Eun-Hee Park – 30
Concluding Remarks
◆
Conclusion - Extension of two-level additive Schwarz preconditioner to the WOPSIP method - Barker, Brenner, Park & Sung, to appear
◆
Current / Future Projects Development of nonoverlapping DD based preconditioners for DG methods including the WOPSIP method: BDDC method, FETI-DP method
Cha-Cha Days 2010
Eun-Hee Park – 31
Concluding Remarks
◆
Conclusion - Extension of two-level additive Schwarz preconditioner to the WOPSIP method - Barker, Brenner, Park & Sung, to appear
◆
Current / Future Projects Development of nonoverlapping DD based preconditioners for DG methods including the WOPSIP method: BDDC method, FETI-DP method
Cha-Cha Days 2010
Eun-Hee Park – 31
Concluding Remarks
◆
Conclusion - Extension of two-level additive Schwarz preconditioner to the WOPSIP method - Barker, Brenner, Park & Sung, to appear
◆
Current / Future Projects Development of nonoverlapping DD based preconditioners for DG methods including the WOPSIP method: BDDC method, FETI-DP method
Cha-Cha Days 2010
Eun-Hee Park – 31
Concluding Remarks
◆
Conclusion - Extension of two-level additive Schwarz preconditioner to the WOPSIP method - Barker, Brenner, Park & Sung, to appear
◆
Current / Future Projects Development of nonoverlapping DD based preconditioners for DG methods including the WOPSIP method: BDDC method, FETI-DP method
Cha-Cha Days 2010
Eun-Hee Park – 31
Concluding Remarks
◆
Conclusion - Extension of two-level additive Schwarz preconditioner to the WOPSIP method - Barker, Brenner, Park & Sung, to appear
◆
Current / Future Projects Development of nonoverlapping DD based preconditioners for DG methods including the WOPSIP method: BDDC method, FETI-DP method
Ω1
Ω2 Γ
Cha-Cha Days 2010
Eun-Hee Park – 32
References
[1] D. N. Arnold, An interior penalty finite element method with discontinuous elements, SIAM J. Numer. Anal., 19 (1982), pp. 742–760. [2] S. C. Brenner, T. Gudi, L. Owens, and L.-Y. Sung, An intrinsically parallel finite element method, J. Sci. Comput., 42 (2010), pp. 118–121. [3] S. C. Brenner, L. Owens, and L.-Y. Sung, A weakly over-penalized symmetric interior penalty method, Electron. Trans. Numer. Anal., 30 (2008), pp. 107–127. [4] M. Dryja and O. B. Widlund, An additive variant of the Schwarz alternating method in the case of many subregions, Technical Report 339, Department of Computer Science, Courant Institute, (1987). [5] J. Douglas, Jr. and T. Dupont, Interior penalty procedures for elliptic and parabolic Galerkin methods, in Lecture Notes in Phys. 58, Springer-Verlag, Berlin, 1976. `re, M. F. Wheeler, and V. Girault, Improved energy estimates for interior [6] B. Rivie penalty, constrained and discontinuous Galerkin methods for elliptic problems I, Comput. Geosci., 3 (1999), pp. 337–360. [7] M. F. Wheeler, An elliptic collocation-finite element method with interior penalties, SIAM J. Numer. Anal., 15 (1978), pp. 152–161.
Cha-Cha Days 2010
Eun-Hee Park – 33
Intrinsic Parallelism of WOPSIP method
a(v, w) =
T ∈Th
T
1 ∇v · ∇w dx + [ v(me )]] · [ w(me )]] |e|2
- Element-wise ordering
e∈Eh
- Edge-wise ordering e
v k+1
v k+2
wk+1
T
wk
vk
Ah w = P t DP w + Jw - D: Block-diagonal matrix where each block is 3 × 3 - J: Block-diagonal matrix where each block is either 1 × 1 or 2 × 2
Cha-Cha Days 2010
Eun-Hee Park – 34