Schwarz-type methods and their application in geomechanics

Transcription

1 Schwarz-type methods and their application in geomechanics R. Blaheta, O. Jakl, K. Krečmer, J. Starý Institute of Geonics AS CR, Ostrava, Czech Republic PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 1/23

2 Contents Introduction Iterative solvers Space decomposition Displacement decomposition Domain decomposition Two-level domain decomposition Parallel implementation Testing environment Academic model problem, results Large-scale real-life problem, results Two-level hybrid preconditioners Testing of solvers, results Present and future work, conclusion PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 2/23

3 Introduction Many physical phenomena can be simulated with the aid of the finite element (FE) method. In many cases, this simulation leads to huge computational demands, which are concentrated mainly to the solution of large-scale systems of linear equations. We shall consider the case, when the investigated physical phenomena are described by elasticity problems on domain Ω R 3. We shall also assume that the FE discretization of these boundary value problems leads to the solution of large-scale systems of the type Au = f, u, f R n, where A is a symmetric positive definite n n matrix. Such systems arise also e.g. from mathematical modelling of stationary diffusion, heat conduction or filtration. PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 3/23

4 Iterative solvers PCG method: GPCG[m] method: r 0 = f Au 0 v 0 = g 0 = G(r 0 ) σ 0 = g 0, r 0 for i = 0, 1,... until r i ε f : w = Av i α = σ i / v i, w u i+1 = u i + αv i r i+1 = r i αw g i+1 = G(r i+1 ) σ i+1 = g i+1, r i+1 β = σ i+1 /σ i r 0 = f Au 0 v 0 = g 0 = G(r 0 ) σ 0 = g 0, r 0 for i = 0, 1,... until r i ε f : w = Av i α = σ i / v i, w u i+1 = u i + αv i r i+1 = r i αw g i+1 = G(r i+1 ) σ i+1 = g i+1, r i+1 v i+1 = g i+1 end v i+1 = g i+1 + βv i for k = min{i + 1, m},..., 1: w = r i+2 k r i+1 k Solvers: sequential parallel data decomposition & parallel operations (matrix by vector multiplication, vector updates,... ) construction of efficient preconditioner & its effective implementation end end βi+1 k = g i+1, w /σ i+1 k v i+1 = v i+1 + βi+1 k v i+1 k Note: In our computations ε = PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 4/23

5 Space decomposition V = V V p, V R n, V k R n k V k, V l... do not need to be linearly independent Restrictions: Prolongations: Decomposition of A: R k : R n R n k R k v 1. v p = vk I k : R n k R n [ 0 I k v k = v k 0 ] A = A 11 A 1p A p1 A pp Algorithm g = G(r): g 0 = 0 for k = 1,..., p: end g k = g k 1 + I k A 1 kk R kz k Type of preconditioning: additive: z k = r multiplicative: z k = (r Ag k 1 ) Solution of A 1 kk : exactly, e.g. by an incomplete factorization inexactly, e.g. by inner (PCG) iterations (up to some lower accuracy, ε = 10 1 ) PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 5/23

6 Displacement decomposition u = (u 1,..., u N ), u i = u ix u iy u iz, V = V x + V y + V z ND = 3 NN {}}{ X 1Y 1Z 1X 2Y 2Z 2... NN ND {}}{ NN NN {}}{{}}{{}}{ X 1X 2... Y 1Y 2... Z 1Z 2... X 1 Y 1 Z 1 X 2 Y 2 Z 2... A X 1 Y 1 Z 1 X 2 Y 2 Z 2... X 1 X 2... Y 1 Y 2... A 11 A 12 A 13 X 1 X 2... Y 1 Y 2... A 21 A 22 A 23 Z 1 Z 2... Z 1 Z 2... stiffness matrix vector A 31 A 32 A 33 stiffness matrix vector v = v 1 v 2 v 3 A = A 11 A 12 A 13 A 21 A 22 A 23 A 31 A 32 A 33 B = A 11 A 22 A 33 This technique naturally allows to split computations just among 3 concurrent tasks. A full scalability of computations is possible only in combination with another SD technique. Blocks in preconditioner are completely independent, therefore no communication of concurrent tasks during the preconditioning is needed. PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 6/23

7 Domain decomposition Ω = p Ω i, V = p i=1 i=1 V i, V i = {v V : v = 0 in Ω\Ω i } Y Z X Ω Ω 1 Ω 2 Ω Ω 1 Ω 2 Ω 3 X 1 Y 1 Z 1 X 2 Y 2 Z 2... ND = 3 NN {}}{ X 1Y 1Z 1X 2Y 2Z 2... A stiffness matrix X 1 Y 1 Z 1 X 2 Y 2 Z 2... vector X 1 Y 1 Z 1 X 2 Y 2 Z 2... X 1Y 1Z 1X 2Y 2Z 2... A 1 A 2 stiffness matrix A 3 X 1 Y 1 Z 1 X 2 Y 2 Z 2... vector An advantage of domain decomposition in only dimension: each of concurrent tasks has to communicate with at most two others corresponding to neighbouring subdomains. Nonoverlapping sds. - block Jacobi: g = G 1 (r) = p i=1 I i à 1 i R i r Overlapping sds. - additive Schwarz: g = G 2 (r) = p i=1 I i A 1 i R i r PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 7/23

8 Two-level domain decomposition Coarse grid: p V = V c + V i, A c, R c, I c i=1 Aggregations: Y Z X Ω Ω c A construction of R c, I c depends on fact, if the coarse grid is nested or non-nested. Two-level additive Schwarz: g = G 3 (r) = G 2 (r) + I c A 1 c R c r An example of regular aggregations. In practice, it is often impossible to make the explicit coarse grid or too much time demanding to make or use R c, I c. In this case, aggregations allow a simple construction of A c, R c, I c. PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 8/23

9 Parallel implementation step k th worker process communication l th worker process data in memory: data in memory: u k, v k, r k, g k, w k w kk, A kk u l, v l, r l, g l, w l w ll, A ll w ki, A ik, i = 1,..., m, i k w li, A il, i = 1,..., m, i l S 1 for i = 1,..., m : w ki = A ik v k for i = 1,..., m : w li = A il v l C 1 = w kl w lk = m m S 2 w k = ik w l = il i=1 i=1 S 3 s k = w k, v k s l = w l, v l m C 2 s = i i=1 S 4 α = s 0 /s α = s 0 /s u k = u k + αv k u l = u l + αv l r k = r k αw k r l = r l αw l S 5, C 3 g = G(r) r k, r l g = G(r) S 6 s k = g k, r k s l = g l, r l m C 4 s 1 = s i i=1 S 7 β = s 1 /s 0 β = s 1 /s 0 s 0 = s 1 s 0 = s 1 v k = g k + βv k v l = g l + βv l The worker-to-worker implementation of one iteration of the parallel CG algorithm with a Schwarz-type preconditioner. PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 9/23

10 Testing environment The parallel implementation of our solvers was realized in Fortran 77 with the aid of PVM, MPI (MPICH or LAM) or PETSc. We used the following parallel computers: Lomond (EPCC Edinburgh): Sun HPC cluster. Front-end: SMP system HPC 3500 with 8 UltraSPARC-II/400 and 8 GB of shared memory. Back-end: 2 SMP system Sunfire 6800 with 24 UltraSPARC-III/750 and 48 GB of shared memory. OS Solaris 2.8, Sun Grid Engine, MPI (Sun), PETSc (2.1.3). Thea (IGAS Ostrava): a Beowulf cluster, file server + interactive node + 8 nodes, each equipped by AMD Athlon/1400, 768 MB of memory, 2 FastEthernet. OS Debian Linux, MPI (MPICH or LAM 6.5.6), PVM (3.4.2), PETSc (2.1.3 and 2.2.0). interactive node computing nodes to Inter/Intranet FastEth. (system) switch FastEth. (appl.) Gigabit file server to Inter/Intranet PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 10/23

11 Academic model problem As a basic test problem, we use the square footing, which is a 3D elasticity problem of soil mechanics. The square area m on the top side of the domain m is loaded by the uniform pressure equal to 2.4 MPa. Due to the symmetry of the considered domain, we work only with its quarter discretized by a rectangular grid with a grid refinement under the footing. Benchmark Used grid # Equations FOOT 40 E FOOT 60 E PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 11/23

12 Lomond - numbers of iterations Test: FOOT 60 E PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 12/23

13 Lomond - times Test: FOOT 60 E PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 13/23

14 Thea - times Test: FOOT 60 E PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 14/23

15 Large-scale real-life problem Another benchmark problem is derived from the real-life large-scale mathematical model of a uranium ore mine Dolní Rožínka (DR) in the Bohemian-Moravian Highlands. The model is considered for the comparison of different mining methods in relation to a development of induced stress fields and a possibility of dangerous rockbursts. Since we presented the procedure and results in a fairly detailed manner already, in this place we will only summarize them. The DR model considers a domain of meters located 800 meters bellow surface. East-West cross-section of the deposit Rožná and the considered area in the left down corner. PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 15/23

16 Dolní Rožínka The domain is discretized by a regular grid of nodes and the resulting linear system has DOF. Z X Y The FE mesh. The whole task simulates four stages of mining, represented by a four-step sequence of problems with different material distributions. For needs to use the big-size testing problem only, we work with the last step of the whole modelling sequence. PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 16/23

17 Results Lomond Thea Solver # P # It T [s] # It T [s] SEQ-IF DiD-IF DiD-II DiD-II* DD-AS DD-AS DD-AS DD-AS DD-AS DD-2LAS DD-2LAS DD-2LAS DD-2LAS DD-2LAS * for the GPCG[1] method PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 17/23

18 Two-level hybrid preconditioners Define: B = p k=1 I k A 1 k R k B c = I c A 1 c R c I k = R T k, I c = R T c Consider the following Schwarz-type preconditioners g = G(r): 1L aditive m. g = Br 2L aditive m. g = Br + B c r 2L nonsymmetric hybrid m. g = Br g = g + B c (r Ag) 2L reverse nonsymmetric hybrid m. g = B c r g = g + B(r Ag) 2L symmetric hybrid m. g = Br g = g + B c (r Ag) g = g + B(r Ag) 2L symmetric averaged hybrid m. g F = Br g C = B c r g F = g F + B(r Ag F ) g C = g C + B c (r Ag C ) g = 1 2 (g F + g C ) 2L symmetric smoothed hybrid m. (with overcorrection η) g = Br g = g + B(r Ag) g = g + ηb c (r Ag) g = g + B(r Ag) g = g + B(r Ag) PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 18/23

19 Thea - numbers of iterations Test: FOOT 40 E Preconditioner: Number of processors type and properties A one-level A two-level H two-level symmetric H nonsymmetric H reverse nonsymmetric H symmetric smoothed without overcorrection H symmetric smoothed with overcorrection η = H symmetric, averaged for GPCG[1] method only PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 19/23

20 Thea - times Test: FOOT 40 E Preconditioner: Number of processors type and properties A one-level A two-level H two-level symmetric % H nonsymmetric H reverse nonsymmetric H symmetric smoothed without overcorrection H symmetric smoothed with overcorrection η = H symmetric, averaged for GPCG[1] method only PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 20/23

21 Thea - number of iterations Test: FOOT 40 E Symmetric hybrid preconditioner with overcorrection: influence of overcorrection η η Number of processors η Number of processors PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 21/23

22 Present and future work Coarse grid with interfaces: Original grid Subdomain 1 Subdomain 2 Aggregations out of interface Coarse grid with interface PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 22/23

23 Conclusion This work outlines parallel implementation of the conjugate gradient method with various preconditioners and shows some applications of the implemented methods. The results prove that such parallel solvers enable efficient solution of large-scale real-life engineering problems even on clusters of common and relatively cheap PC s. Acknowledgment: Authors are grateful for the support of the Academy of Sceinces of the Czech Republic, contract No. S , and the Grant Agency of the Czech Republic, contract No. 105/04/P036. PDEMAMIP, September 7-11, 2004,Sunny Beach, Bulgaria R. Blaheta, O. Jakl, K. Krečmer, J. Starý: Schwarz-type methods and their application in geomechanics page 23/23