Domain decomposition strategies with black box subdomain solvers

of 18 /18
Domain decomposition strategies with black box subdomain solvers Silvia Bertoluzza Istituto di Matematica applicata e Tecnologie Informatiche del CNR, Pavia Grenoble, May 19th, 2011 Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 1 / 18

Embed Size (px)

description

Presentation on Domain decomposition strategies with black boxsubdomain solvers

Transcript of Domain decomposition strategies with black box subdomain solvers

  • Domain decomposition strategies with black boxsubdomain solvers

    Silvia Bertoluzza

    Istituto di Matematica applicata e Tecnologie Informatiche del CNR, Pavia

    Grenoble, May 19th, 2011

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 1 / 18

  • Motivations

    Domain decomposition

    Solution of differential problem on = k reduced to solvingproblems in the k s + coupling condition

    Many advantages:

    PreconditioningParallelizationPossibility of treating bigger problem: divide et impera

    Two classes

    Conforming: continuity is imposed strongly across the subdomain edgesNon conforming: continuity is imposed weakly across the subdomainedges

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 2 / 18

  • Motivations

    Non conforming methods

    Allows to couple discretizations of different type:

    FE + spectral methodswavelets + FEFE + FE on non matching grids

    Flexibility in the choice of meshes / Adaptivity

    Possibility of using the locally best method

    Allows to use different models in different subdomain

    Different approaches: Mortar method, 3 Fields formulation, Lagrangemultiplier approach, . . .

    Drawback: coupling condition saddle point problem: stabilityproblems and limitations due to inf-sup conditions

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 3 / 18

  • Motivations

    Bulk of the computation: solution of local problems

    Great effort spent on optimization of mono-domain codes for solvingPDEs

    Two level of optimization

    Algorithm optimization (minimize computational cost)Architecture dependent optimization at compiling time

    Goal

    Develop domain decomposition methods allowing the use of optimizedcodes as subdomain solvers in a (non conforming) domain decompositionmethod

    Important: black-box SD solver

    Do not want to touch the SD code in any way

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 4 / 18

  • Motivations

    Use black box SD solvers

    Many problems to be faced

    Informatics problems: need to be able two make two differentinstances of a code (or two different codes!) talk to each other. Hugeproblem

    Mathematical problems

    design global solution strategy using standard local solversgive conditions ensuring convergence of the iterative strategydesign suitable preconditionersneed to be able to verify that the SD solvers verify the conditions forconvergence

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 5 / 18

  • Motivations

    To fix the ideas

    Model problem{u = f in u = 0 on

    Domain decomposition

    u = f in k , k[u] = 0 on

    [u] = 0 on

    + BC ([] =jump)

    Geometry

    = kkk = k \ k = k

    3

    1

    1

    3

    2 2

    3

    1

    1

    3

    2 2

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 6 / 18

  • Motivations

    The Steklov-Poincare approach

    Observation: if we knew u on , Dirichlet problems on thesubdomains would give us u on all the k s!

    h: guess for u| Solver uk(h) (guess for u on k)

    Is h a good guess? Check r = [u(h)]. If r 0 then guess isgood!

    Steklov-Poincare based methods: find h minimizing |r |

    Basis for many DD methods

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 7 / 18

  • The FETI approach

    What can codes do?

    Steklov-Poincare operator: guess u| residual: [u|]Problem: computation of of solution not always available

    Input: what are we free to prescribe

    differential operatorr.h.s. f(non homogeneous) Dirichlet boundary conditions(non homogeneous) Neumann boundary conditions

    Output: what can we expect to retrievevalue of the solution at (boundary) nodesouter normal derivative: more rare!

    To be as general as possible design a domain decomposition strategyusing Neumann solvers and imposing continuity of the solution.

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 8 / 18

  • The FETI approach

    Idea: switch the role of trace and normal derivative

    guess = u|solve Neumann problems on subdomains

    k

    u v =

    k

    f v +

    k

    (k)v

    (k = 1 depends on whether is the outer or inner normal w.r. to k) residual: r = [u]| with u solution of Neumann problems

    minimize |r | by some iterative technique (Ex: CG method)

    floating domains solution determined modulo a constantuse FETI approach saddle point problem with a scalar multiplier perfloating domain

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 9 / 18

  • The FETI approach

    Discretization

    Galerkin discretization with h L2() finite dimensional

    Discrete problem is well posed provided h 0h0h = { L2() : `, k |k` = constant}

    use any suitable methodProjected Conjugate Gradient methodConjugate Gradient method on a reduced space

    Replace solution of Neumann problem with black box call to a PDEnumerical solver A+

    Skip analysis

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 10 / 18

  • The FETI approach

    Assumptions

    The discrete solver A+ satisfy the following estimate:

    (A+ A+)f 1, . sf s1

    (weaker assumption) The discrete solver A+h satisfy the followingestimate:

    B(A+h A+)f 1/2, . sf s1

    The space h verify the following inverse inequality

    hs, . hsh1/2,

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 11 / 18

  • The FETI approach

    Theorem

    Assumptions

    The discrete solver A+ satisfy the following estimate:

    (A+ A+)f 1, . sf s1The space h verify the following inverse inequality

    hs, . hsh1/2,the meshsizes for the subdomain problems are sufficiently finer thanthe meshsize of h.

    Then the matrix S corresponding to the discrete problem is symmetricpositive definite, provided. Moreover we have that

    h1/2, . (hs + s)us+1S.B. 2011

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 12 / 18

  • The FETI approach

    Numerical experiments: 2D problem

    Discretization for : piecewise constants on a uniform mesh

    Sobdomain solvers: Matlab PDE toolbox (functions used: meshinit,assempde)

    Test: true solution u = (x + 1)2 + y sin(x + 1) on = (1, 2) (0, 1)Boundary conditions: homogeneous Dirichlet on x = 1,non-homogeneous Neuman elsewhere

    -1 0 2

    1

    -1 -0.5 0 0.5 1 1.5 20

    0.1

    0.2

    0.3

    0.4

    0.5

    0.6

    0.7

    0.8

    0.9

    1

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 13 / 18

  • The FETI approach

    h = 1/N meshsize for h

    = 1/M1 meshsize on 1

    = 1/M2 meshsize on 2

    5 situations

    (a) M1 = M2 = N 1,(b) M1 = M2 = N,(c) M1 = M2 = N + 1,(d) M1 = N + 1, M2 = N 1(e) M1 = 2N, M2 = [1.5N]

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 14 / 18

  • The FETI approach

    N Case (a) Case (b) Case (c) Case (d) Case (e)

    5 - .10093 .08280 .10539 .06271110 - .05002 .050023 .05043 .0288520 - .02416 .02416 .02459 .0143740 - .01201 .01201 .01198 .00718

    Table: Errors for u

    N Case (a) Case (b) Case (c) Case (d) Case (e)

    5 - .04082 .03013 .05866 .0203410 - .01747 .01081 .03873 .0075920 - .01756 .00839 .04806 .0033440 - .00470 .00350 .07849 .00145

    Table: Errors for

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 15 / 18

  • The FETI approach

    Numerical experiment 3D

    Discretization for : cubic splines

    Sobdomain solvers: Comsol+Matlab (functions used: )

    Right hand side f = x + 3 y on = (1, 2) (0, 1) (0, 1)Boundary conditions: homogeneous Dirichlet on x = 1,homogeneous Neumann elsewhere

    Comsol: Domain and Grid Comsol: Soluzione

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 16 / 18

  • The FETI approach

    Convergence of the CG algorithm

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 17 / 18

  • The FETI approach

    Conclusions

    Different strategies available to use standard solvers in domaindecomposition

    Possibility of using Dirichlet or Neumann solver

    Inf-sup or inf-sup type conditions need to be satisfied

    Stabilization techniques that do not touch the solvers are available

    Open issues

    Preconditioning (easy)Monitoring the validity of the assumptionsExtending to other class of problems

    Thank you!!

    Silvia Bertoluzza (IMATI-CNR) DD with Black Box SD solvers Grenoble, May 19th, 2011 18 / 18

    MotivationsThe FETI approach