fmm accelerated bem for 3d laplace & helmholtz...
TRANSCRIPT
FMM accelerated BEMfor 3D Laplace & Helmholtz equations
Nail A. Gumerov and Ramani DuraiswamiInstitute for Advanced Computer Studies
University of Maryland, U.S.A.
Presentation for the International Conference on Boundary Element Technique BETEQ-7, September 2006, Paris, France
© Gumerov & Duraiswami, 2006
Outline
• Introduction• Problem formulation• BEM speed up with the FMM• Performance tests
• Laplace equation• Helmholtz equation
• Conclusions
© Gumerov & Duraiswami, 2006
Introduction
• BEM is a great method
• Drawbacks of the standard method• Slow: O(N3)-direct, O(NiterN2)-iterative• Memory expensive: O(N2)-full matrix storage• For modern PC’s are not suitable for solution
of large problems, N > 104
• FMM acceleration/memory reduction• Fast: O(NiterN) (or O(NiterNlogN))• Memory inexpensive: O(N) (or O(NlogN))• Modern PC’s can solve large problems, N ~ 106
© Gumerov & Duraiswami, 2006
Large problems: Example 1Introduction
© Gumerov & Duraiswami, 2006
Mesh: 30000 elements15002 vertices
Acoustic scattering from a sphere(model of human head)
Mesh: 33082 elements16543 vertices
Large problems: Example 1/1Introduction
© Gumerov & Duraiswami, 2006
KEMAR: 405002 elements, 202503 vertices
Large problems: Example 2Introduction
© Gumerov & Duraiswami, 2006
Multiparticle system (e.g. blood flow)
Large mesh (hundreds of thousands elements)
Problem formulation
© Gumerov & Duraiswami, 2006
Laplace: Helmholtz:
Boundary conditions:
For external problems: For external problems:
where
n
S
Layer potentials
© Gumerov & Duraiswami, 2006
Laplace: Helmholtz:
Problem formulation
Normal derivatives:
Boundary integral equations
© Gumerov & Duraiswami, 2006
Problem formulation
Green’s identity (direct) formulation:
Layer potential (indirect) formulation:
BEM speed up with the FMM
© Gumerov & Duraiswami, 2006
• Use iterative methods (we use GMRES)• The system matrix can be decomposed as
A = Asparse+Adense .• The sparse part includes diagonal
elements+neighbor element influence• Integrals can be computed using standard BEM
techniques, O(N) memory/operation complexity• The dense part is related to far element
influence• Most memory/computationally expensive• Can be done with the FMM• Low order quadratures over elements can be used
BEM and FMM data structures
© Gumerov & Duraiswami, 2006
BEM speed up with the FMM
FMM neighborhood
BEM neighborhood
This element is close to red one in all senses(high order quadrature, influence directly)
This element is far to red one in all senses(low order quadrature, influence via expansion)
This element is far to red one in the sense of BEM, but close in the sense of the FMM(low order quadrature, influence directly)
rmin
Level lmax
Computation of normal derivatives
© Gumerov & Duraiswami, 2006
BEM speed up with the FMM
Differential operators in the space of expansion coefficients:
Detailed theory and expressions can be found in bookN.A. Gumerov and R. Duraiswami, Fast Multipole Methods for the Helmholtz Equation in Three Dimensions, Elsevier, Oxford UK, 2004.
sparse matrix
Expansion of boundary integrals
© Gumerov & Duraiswami, 2006
BEM speed up with the FMM
Greens function expansion:
Quadratures:
Expansion coefficients:
Particulars of the FMM employed
© Gumerov & Duraiswami, 2006
BEM speed up with the FMM
• Two or four BEM matrix-vector multiplications are performed at “one shot” using sparse matrix representation of differential operators
• O(p3) translation methods based on the RCR*
-decomposition are used both for the Laplace and Helmholtz equation
• Details can be found in our technical reports and FMM papers, 2003-2006.
* RCR-decomposition is the Rotation-Coaxial translation-Rotation decomposition of an arbitrary translation operator in three dimensions
Basic FMM flow chart
© Gumerov & Duraiswami, 2006
BEM speed up with the FMM
1. Get S-expansion coefficients(directly)
2. Get S-expansion coefficients from children
(S|S translation)
Level lmax
3. Get R-expansion coefficients from far neighbors
(S|R translation)
Level 2
4. Get R-expansion coefficients from far neighbors
(S|R translation)
6. Evaluate R-expansions (directly)
7. Sum sourcesin close neighborhood
(directly)
Start
End
Level lmaxLevel lmax
5. Get R-expansion coefficients from parents
(R|R translation)
Levels lmax-1,…, 2 Levels 3,…,lmax
1. Get S-expansion coefficients(directly)
2. Get S-expansion coefficients from children
(S|S translation)
Level lmax
3. Get R-expansion coefficients from far neighbors
(S|R translation)
Level 2
4. Get R-expansion coefficients from far neighbors
(S|R translation)
6. Evaluate R-expansions (directly)
7. Sum sourcesin close neighborhood
(directly)
Start
End
Level lmaxLevel lmax
5. Get R-expansion coefficients from parents
(R|R translation)
Levels lmax-1,…, 2 Levels 3,…,lmax
Performance tests
© Gumerov & Duraiswami, 2006
• Flat triangular surface discretization• Accuracy of the FMM selected to be
consistent with the accuracy of the BEM (relative errors normaly within the range 10-2 -10-4)
• Problem sizes 500-3,000,000 boundary elements
• Where possible compared with direct BEM, iterative BEM, and low memory direct BEM
• Hardware: PC: 3.2 GHz, Intel Xeon processor, 3.5 GB RAM
Laplace equation
© Gumerov & Duraiswami, 2006
Performance tests
1000 randomly oriented ellipsoids488000 vertices and 972000 elements
Potential
Typical test configuration: for external Dirichletand Neumann problems
Laplace equation
© Gumerov & Duraiswami, 2006
Performance tests
1.E-02
1.E-01
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+02 1.E+03 1.E+04 1.E+05 1.E+06Number of Vertices, N
Tota
l CP
U ti
me
(s)
GMRES+FMM
GMRES+Low Mem Direct
GMRES+High Mem Direct
LU-decomposition
O(N2) Memory Threshold
y=ax
y=bx2
y=cx3
BEMDirichlet Problemfor Ellipsoids(3D Laplace)
FMM: p = 8
Laplace equation
© Gumerov & Duraiswami, 2006
Performance tests
1.E-03
1.E-02
1.E-01
1.E+00
1.E+01
1.E+02
1.E+02 1.E+03 1.E+04 1.E+05 1.E+06Number of Vertices, N
Sing
le M
atrix
-Vec
tor M
ultip
licat
ion
CPU
Tim
e (s
)
y=ax
y=bx2Number of GMRES Iterations
FMM
Direct+Matrix EntriesComputation
Multiplicationof Stored Matrix
3D Laplace
FMM: p = 8
Helmholtz equation
© Gumerov & Duraiswami, 2006
Performance tests
Incident wavePlane wave scattering from a sphere, ka=30BEM/FMM, 240002 vertices, 480000 elements
(comparison with analytical solution)
GMRES: 115 iterations, ε =10-5,FMM: lmax= 6, pmax= 22
Helmholtz equation
© Gumerov & Duraiswami, 2006
Performance tests
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+03 1.E+04 1.E+05 1.E+06Number of Vertices, N
Tota
l CP
U ti
me
(s)
GMRES+FMM
GMRES+Low Mem Direct
GMRES+High Mem Direct
LU-decomposition
O(N2) Memory Threshold
y=ax
y=bx2
y=cx3
BEMNeumann Problemfor Sphere (3D Helmholtz)
kD=34.64
Helmholtz equation
© Gumerov & Duraiswami, 2006
Performance tests
1.E-02
1.E-01
1.E+00
1.E+01
1.E+02
1.E+03
1.E+03 1.E+04 1.E+05 1.E+06Number of Vertices, N
Sing
le M
atrix
-Vec
tor M
ultip
licat
ion
CPU
Tim
e (s
)
y=ax
y=bx2Number of GMRES Iterations
FMM
Low MemDirect
Multiplicationof Stored Matrix
3D Helmholtz, kD=33.64
Helmholtz equation
© Gumerov & Duraiswami, 2006
Performance tests
Mesh: 249856 vertices/497664 elements
kD=29, Neumann problem kD=144, Robin problem(impedance, sigma=1)
(some other scattering problems were solved)
Helmholtz equation
© Gumerov & Duraiswami, 2006
Performance tests
(some other scattering problems were solved)
Incident plane wave
Mesh: 132072 elements 65539 vertices
exp(ikz)
Helmholtz equation
© Gumerov & Duraiswami, 2006
Performance tests
(some other scattering problems were solved)
kD=0.96 kD=9.6 kD=96(250 Hz)
Sound pressure
(2.5 kHz) (25 kHz)
Conclusions
© Gumerov & Duraiswami, 2006
• Developed method (BEM with GMRES+FMM) shows a good performance and can be used for solution of large boundary value problems for the Laplace and Helmholtz equtaions (up to million size) on a single PC
• Future research: • Parallel versions of the method should bring additional
speedups• BEM for different equations (e.g. biharmonic and Stokes
equations) can be also speeded up in the same way • Efficient iterative methods for BEM are of high demand,
as the matrix-vector products can be accelerated via the FMM
Software Availability:
© Gumerov & Duraiswami, 2006
• FMM-based software is licensed from the University of Maryland to Fantalgo, LLC
• Includes matrix-vector product for 3D Laplace, biharmonic, Helmholtz (Maxwell is coming) equations and Gaussian kernel
• BEM and function fitting via iterative solvers• Contact
[email protected] further information
THANK YOU !