Incomplete Cholesky preconditioners that exploit the low-rank property

Transcription

1 ; anapov/ 1 / 35 Incomplete Cholesky preconditioners that exploit the low-rank property (theory and practice) Artem Napov Service de Métrologie Nucléaire, Université Libre de Bruxelles PRECONDITIONING (TU EINDHOVEN) JUNE 17, 2015

2 Motivation The solution of a symmetric positive definite (SPD) system Au = b may be obtained with a direct method (usually: Cholesky factorization A = R T R) robust iterative method (usually: preconditioned conjugate gradient) ; efficient if a good SPD preconditioner B A is available, i.e.: 1. Cheap (to construct, store, invert, parallelize) 2. Good approximation of A contradictory goals tradeoff Possible black box design: take an exact factorization (here, Cholesky) ; add approximation (to enforce 1.) incomplete factorization preconditioner B = R T R A. 2 / 35

3 Motivation Incomplete factorizations B = R T R A may perform approximations by dropping individual entries (mainstream) ; using low-rank approximations (emerging). Incomplete factorizations: + are (almost) black box (approximation threshold required) + are (relatively) robust may breakdown (but breakdown-free variants exist) have no guarantee to converge fast; fast convergence if B approximates well A ; that is, if κ(r T AR 1 ) = λ max(r T AR 1 ) λ min (R T AR 1 ) is small. Here we present incomplete factorizations that are breakdown-free and have controllable condition number. 3 / 35

4 4 / 35 Outline Theory... One-level variant basic approach and underlying analysis Multilevel variants extensions of analysis to multi-level setting... and practice Sparse solver motivation and design choices Numerical experiments and comparison with other solvers

5 4 / 35 Outline Theory... One-level variant basic approach and underlying analysis Multilevel variants extensions of analysis to multi-level setting... and practice Sparse solver motivation and design choices Numerical experiments and comparison with other solvers

6 5 / 35 One-level variant : basic ideas Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) A ( k = 1 )

7 5 / 35 One-level variant : basic ideas Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) ( k = 1 ) A 11 A 12 A 22 A

8 5 / 35 One-level variant : basic ideas Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) ( k = 1 ) R 11 A 12 A 22 A 11 = R T 11 R 11 A

9 5 / 35 One-level variant : basic ideas Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) ( k = 1 ) R 11 R T 11 A 12 A 22 A

10 5 / 35 One-level variant : basic ideas Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) ( k = 1 ) R 12 A 22 A R 12 R 12 The dropping is orthogonal if R ( ) T R 12 R 12 R R 12 = O. R 12 R 12

11 5 / 35 One-level variant : basic ideas Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) ( k = 1 ) R 12 A 22 A R 12 R 12 The dropping is orthogonal if R 12 T ( R 12 R 12 ) = O. This implies monotonicity for Schur complement (for any SPD A!) : v T (A 22 R T 12 R 12 )v v T (A 22 R T 12R 12 )v

12 One-level variant : basic ideas Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) ( k = 1 ) R 12 A 22 A R 12 R 12 Example of orthogonal dropping low-rank approximation via truncated SVD (with absolute threshold tol a ) ) T + U 2 (Σ 2 V2 ) }{{} R 12 = ( U 1 U 2 ) ( Σ 1 Σ 2 ) ( V T 1 V T 2 If Σ 2 < tol a, then = U 1 (Σ 1 V T 1 ) }{{} R 12 R 12 R 12 R 12 R 12 < tol a 5 / 35

13 5 / 35 One-level variant : basic ideas Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) R 12 A 22 B 1 ( k = 1 ) S 1 = A 22 1 R T 12 1 R 12

14 5 / 35 One-level variant : basic ideas Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) B 1 ( k = 2 )

15 5 / 35 One-level variant : basic ideas Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) B 1 ( k = 2 )

16 One-level variant : successive approximations Incomplete (block left-looking) Cholesky for each block row k = 1,..., l : 1 factorize 2 solve 3 approximate 4 update (compute new rows) ( k = 2 ) approximate new rows only One-level (model) variant k = 1 k = 2 k = 3 k = 4 6 / 35

17 7 / 35 Accuracy of individual dropping The orthogonal dropping is assumed, namely R 12 T ( R 12 R 12 ) = O R 12 A 22 B k 1 The accuracy of individual dropping at step k : (R γ k = 12 R ) 12 S 1/2 < 1, B with S B = A 22 R T 12 R 12. R 12 B k

18 Model problem Model Problem: { u u = f in Ω = (0, 1) 2 u n Full system matrix: = 0 on Ω A Ω = A I A II A I,Γ A II,Γ. Ω I Γ Ω II A T I,Γ AT II,Γ A Γ Interface system matrix: A = A Γ A T I,ΓA 1 I A Γ,I A T II,ΓA 1 II A Γ,II Corresponds to: last Schur complement in sparse Cholesky with ND ordering Schur complement system in iterative substructuring 8 / 35

19 9 / 35 Accuracy of individual dropping The orthogonal dropping is assumed, namely R 12 T ( R 12 R 12 ) = O. The accuracy of individual dropping at step k: (R γ k = 12 R ) 12 S 1/2 < 1, with B S B = A 22 R T 12 R 12. Model problem: (k = 1, block size = 10, grid, truncated SVD dropping) r γ B k 1 B k

20 One-level variant: conditioning analysis (for λ max ) Assume A SPD and that dropping is orthogonal Let B k correspond to the preconditioner at the end of step k, with A = B 0 B l = R T R is the final preconditioner. Only newly computed rows are modified; Let λ (k) ( ) max = λ max B 1 (k 1) l B k, λ max = λ max ( B 1 l B k 1 ) B k 1 B k Then λ (k) max λ (k 1) max λ (k) max + g( λ (k) max, γ k ) where g( λ, γ ) = max β>0 2 γ β λ 1 β 2 β 2 + λ 1. B l 10 / 35

21 One-level variant: conditioning analysis (for λ min ) Assume A SPD and that dropping is orthogonal Let B k correspond to the preconditioner at the end of step k, with A = B 0 B l = R T R is the final preconditioner. Only newly computed rows are modified; Let λ (k) min = λ ( ) min B 1 (k) l B k, λ min = λ ( ) min B 1 l B k 1 B k 1 B k Then λ (k) min λ(k 1) min λ (k) min g( λ(k) min, γ k ) where g( λ, γ ) = max β>0 2 γ β λ 1 β 2 β 2 + λ 1. B l 10 / 35

22 One-level variant: conditioning analysis One-level bounds (for λ max ) λ (k) max λ (k 1) max λ (k) max + g( λ (k) max, γ k ) B k 1 where λ (k) max ( ) = λ max B 1 (k 1) l B k, λ max = λ max ( B 1 l B k 1 ) The estimate is sharp: For any γ k a matrix A < 1, k = 1,..., l, there exist a sequence of approximations B k (with B 0 = A) such that the bounds for λ (0) max and λ (0) min are simultaneously reached. B k B l 11 / 35

23 Model problem: numerical experiments Model Problem: { u u = f in Ω = (0, 1) 2 u n Full system matrix: = 0 on Ω A Ω = A I A II A I,Γ A II,Γ. Ω I Γ Ω II N A T I,Γ AT II,Γ A Γ Interface system N N matrix: A = A Γ A T I,ΓA 1 I A Γ,I A T II,ΓA 1 II A Γ,II Algorithmic details: block size = 10 we consider N from 20 to 650 (n from 400 to , l from 1 to 64) maximal ranks remains below 4 12 / 35

24 13 / 35 One-level variant: numerical experiments (tol r = 10 3 ) 10 2 κ(r T AR 1 ) n condition number one-level precond. upper bound one-level

25 14 / 35 Outline Theory... One-level variant basic approach and underlying analysis Multilevel variants extensions of analysis to multi-level setting... and practice Sparse solver motivation and design choices Numerical experiments and comparison with other solvers

26 Multilevel variants : motivation One-level (model) variant k = 1 k = 2 k = 3 k = 4 O(N 3 ) cost, O(N 2 ) memory (although both cost and memory are improved!) N 15 / 35

27 Multilevel variants : motivation General \ Sequentially Semiseparable variant (SSS) [Gu, Li, Vassilevski 10] k = 1 k = 2 k = 3 k = 4 O(rN 2 ) cost, O(rN) memory r - maximal approximation rank N N/r improvement! 16 / 35

28 17 / 35 Multilevel variants : general estimate B 1 Assuming that dropping is orthogonal one has ( ) λ max B 1 k B k 1 = 1 + γk ( ) λ min B 1 k B k 1 = 1 γk B 2 and, hence, κ(r T AR 1 ) l k=1 1 + γ k 1 γ k. B 3 B 4

29 Model problem Model Problem: { u u = f in Ω = (0, 1) 2 u n Full system matrix: = 0 on Ω A Ω = A I A II A I,Γ A II,Γ. Ω I Γ Ω II N A T I,Γ AT II,Γ A Γ Interface system N N matrix: A = A Γ A T I,ΓA 1 I A Γ,I A T II,ΓA 1 II A Γ,II Algorithmic details: block size = 10 we consider N from 20 to 650 (n from 400 to , l from 1 to 64) maximal ranks remains below 4 18 / 35

30 19 / 35 Multilevel variants: numerical experiments κ(r T AR 1 ) n condition number κ(a) SSS precond. (tol r = 10 3 ) upper bound general (10 3 )

31 20 / 35 Multilevel variants: nested subspaces Sequentially Semiseparable variant (SSS) [Gu, Li, Vassilevski 10] k = 1 k = 2 k = 3 k = 4 Nested subspaces assumption R 12 R 12 R 12 R 12 B 1 B 2 B 3 B 4 span( R 12 ) span( ) span( ) span( )

32 Multilevel variants: nested subspaces Sequentially Semiseparable variant (SSS) [Gu, Li, Vassilevski 10] k = 1 k = 2 k = 3 k = 4 Nested subspaces assumption R 12 R 12 R 12 R 12 B 1 B 2 B 3 B 4 span( R 12 ) span( R 12 ) span( ) span( ) 20 / 35

33 Multilevel variants: nested subspaces Sequentially Semiseparable variant (SSS) [Gu, Li, Vassilevski 10] k = 1 k = 2 k = 3 k = 4 Nested subspaces assumption R 12 R 12 R 12 R 12 B 1 B 2 B 3 B 4 span( R 12 ) span( R 12 ) span( R 12 ) span( ) 20 / 35

34 20 / 35 Multilevel variants: nested subspaces Same bounds as for one-level case! One-level bounds (for λ max ) where λ (k) max λ (k) max λ (k 1) max ( ) = λ max B 1 (k 1) l B k, λ max Nested subspaces assumption λ (k) max + g( λ (k) max, γ k ) = λ max ( B 1 l B k 1 ) R 12 R 12 R 12 R 12 B 1 B 2 B 3 B 4 span( R 12 ) span( R 12 ) span( R 12 ) span( )

35 20 / 35 Multilevel variants: nested subspaces Same bounds as for one-level case! One-level bounds (for λ min ) where λ (k) min λ (k) min = λ min λ(k 1) min ( ) B 1 (k 1) l B k, λ min Nested subspaces assumption λ (k) min g( λ(k) min, γ k ) = λ min ( B 1 l B k 1 ) R 12 R 12 R 12 R 12 B 1 B 2 B 3 B 4 span( R 12 ) span( R 12 ) span( R 12 ) span( )

36 21 / 35 Multilevel variants: numerical experiments (tol r = 10 3 ) κ(r T AR 1 ) n condition number SSS precond. one-level precond. upper bound one-level one-level

37 22 / 35 Outline Theory... One-level variant basic approach and underlying analysis Multilevel variants extensions of analysis to multi-level setting... and practice Sparse solver motivation and design choices Numerical experiments and comparison with other solvers

38 23 / 35 Sparse solver: factorization MAIN FEATURES: symmetric left-looking block factorization accumulate updates before block row computation nested dissection (ND) ordering induces block structure, enforces sparsity, and reduces operation count orthogonal low-rank approximations by truncated SVD or rank-revealing QR

39 23 / 35 Sparse solver: factorization MAIN FEATURES: symmetric left-looking block factorization accumulate updates before block row computation nested dissection (ND) ordering induces block structure, enforces sparsity, and reduces operation count orthogonal low-rank approximations by truncated SVD or rank-revealing QR Γ 3 Γ 2 Ω 2 Ω 3 Ω 4 Ω 1 Γ 1 } } } } } } Ω 1 Ω 2 Γ 1 Ω 3 Ω 4 Γ 2 Γ 3

40 23 / 35 Sparse solver: factorization MAIN FEATURES: symmetric left-looking block factorization accumulate updates before block row computation nested dissection (ND) ordering induces block structure, enforces sparsity, and reduces operation count orthogonal low-rank approximations by truncated SVD or rank-revealing QR

41 24 / 35 Sparse solver: factorization SPECIAL FEATURES: rank-revealing reordering inside ND blocks ensures the algebraic character of the solver symbolic compression ensures that memory usage decreases adaptive block size faster than fixed block size

42 25 / 35 Sparse solver: revealing low-rank Model problem analysis [Chandrasekaran et al., 10]: rank is bounded proportionally to the number of connections between the corresp. subset and the remaining nodes of separator Implemented: recursive edge bisection (via METIS) 2D model separators:

43 Sparse solver: revealing low-rank Model problem analysis [Chandrasekaran et al., 10]: rank is bounded proportionally to the number of connections between the corresp. subset and the remaining nodes of separator Implemented: recursive edge bisection (via METIS) 2D model separators: 25 / 35

44 Sparse solver: revealing low-rank Model problem analysis [Chandrasekaran et al., 10]: rank is bounded proportionally to the number of connections between the corresp. subset and the remaining nodes of separator Implemented: recursive edge bisection (via METIS) 2D model separators: } } 25 / 35

45 Sparse solver: revealing low-rank Model problem analysis [Chandrasekaran et al., 10]: rank is bounded proportionally to the number of connections between the corresp. subset and the remaining nodes of separator Implemented: recursive edge bisection (via METIS) 2D model separators: } } } } } } 25 / 35

46 Sparse solver: revealing low-rank Model problem analysis [Chandrasekaran et al., 10]: rank is bounded proportionally to the number of connections between the corresp. subset and the remaining nodes of separator Implemented: recursive edge bisection (via METIS) 2D separators from Scotch: connections of length 2 25 / 35

47 Sparse solver: revealing low-rank Model problem analysis [Chandrasekaran et al., 10]: rank is bounded proportionally to the number of connections between the corresp. subset and the remaining nodes of separator Implemented: recursive edge bisection (via METIS) 2D separators from Scotch: connections of length 2 25 / 35

48 Sparse solver: revealing low-rank Model problem analysis [Chandrasekaran et al., 10]: rank is bounded proportionally to the number of connections between the corresp. subset and the remaining nodes of separator Implemented: recursive edge bisection (via METIS) 3D model separators: 25 / 35

49 Sparse solver: revealing low-rank Model problem analysis [Chandrasekaran et al., 10]: rank is bounded proportionally to the number of connections between the corresp. subset and the remaining nodes of separator Implemented: recursive edge bisection (via METIS) 3D model separators: } } } } } } 25 / 35

50 26 / 35 Sparse solver: symbolic compression If for a low-rank approximation R 12 R 12 Q one has ( nnz R 12 ) nnz( R 12 ) + nnz( Q ), then the low-rank approximation is replaced by R 12 = I R 12. This low-rank approximation is (trivially) orthogonal. No need to store I does not increase memory use.

51 27 / 35 Sparse solver: adaptive block size Via symbolic compression (s.c.), block subdivision can evolve: } } } } }s.c. }s.c. } } } } We use this evolution: by automatically applying s.c. after few occurrences at the same level; (more successive s.c. occur, (exponentially) more s.c. are skipped to save computations); to adjust the minimal block size for the following separators (this also saves computations).

52 28 / 35 Outline Theory... One-level variant basic approach and underlying analysis Multilevel variants extensions of analysis to multi-level setting... and practice Sparse solver motivation and design choices Numerical experiments and comparison with other solvers

53 Numerical experiments: setting Solver parameters: the preconditioner, denoted by SIC, is used with approximation based on rank-revealing QR with column pivoting and absolute approximation threshold ; adaptive block size (initially block size is set to 16); PCG as outer iteration. Test problems: all SPD matrices from University of Florida Sparse Matrix Collection 1 with n > 10 5 and random rhs (excluded: thermomech_tk and bmw7st_1, both have κ(a) ). Experimental setting: time reported for best threshold value (chosen among {10 10, 10 9,..., }) stopping criterion : 10 6 relative residual decrease; hardware : Intel Xeon L5420, 2.5 GHz, 16 GB RAM / 35

54 30 / 35 Comparison with (exact) Cholesky 100 chol SIC time (sec.) per million nnz problems by nnz

55 31 / 35 Comparison with (unpreconditioned) CG 100 CG SIC time (sec.) per million nnz problems by nnz

56 32 / 35 Comparison with ILUPACK ILUPACK SIC time (sec.) per million nnz problems by nnz

57 33 / 35 Conclusions In theory... We have presented a conditioning analysis for incomplete Cholesky factorizations preconditioner based on orthogonal dropping Only requirement on A: should be SPD The analysis relates the condition number of the preconditioned system to the individual dropping accuracies (namely, to γ k s) The analysis is sharp in the one-level case One-level bound can be extended to the multilevel setting if an additional nested subspace assumption holds; it naturally holds for the presented preconditioner

58 34 / 35 Conclusions... and in practice We have presented a preliminary implementation of incomplete Cholesky factorizations preconditioner based on orthogonal dropping The solver targets at sparse matrices; it uses sparsity structure to reduce the block rank during the factorization Preliminary numerical experiments demonstrate that the solver is competitive

59 Further details Theory: A. Napov, Conditioning analysis of incomplete Cholesky factorizations with orthogonal dropping, SIAM J. Matrix Anal. Appl. A. Napov, Conditioning analysis of incomplete Cholesky factorizations with orthogonal dropping II: nested subspaces (in preparation) Related approaches (for dense matrices): M. Gu, X. S. Li, P. Vassilevski, Direction-preserving and Schur-monotonic semiseparable approximations of symmetric positive definite matrices, SIAM J. Matrix Anal. Appl. J. Xia, M. Gu, Robust approximate Cholesky factorization of rank-structured symmetric positive definite matrices, SIAM J. Matrix Anal. Appl. Rank revealing based on sparsity: A. Napov, X. S. Li, An algebraic multifrontal preconditioner that exploits the low-rank property, Numer. Lin. Alg. Appl. (to appear). 35 / 35