Linear and Nonlinear SP 2 Methods for Large Scale Eigenvalue - PowerPoint PPT Presentation

Linear and Nonlinear “ SP 2 ” Methods for Large Scale Eigenvalue Calculations Zhaojun Bai University of California, Davis Workshop on Recent Advances in Numerical Methods for Eigenvalue Problems NCTS, Taiwan, January 4-8, 2008

Linear and Nonlinear “ SP 2 ” Methods for Large Scale Eigenvalue Calculations (SP 2 = Structure-Preserving Subspace Projection) Zhaojun Bai University of California, Davis Workshop on Recent Advances in Numerical Methods for Eigenvalue Problems NCTS, Taiwan, January 4-8, 2008

Two parts of this talk • Part I: Linear problems Algebraic Substructuring for eigenvalue and frequency response calculations • Part II: Nonlinear problems Nonlinear Rayleigh-Ritz ITerative (NRRIT) technique for solving nonlinear eigenvalue problems

Part I Algebraic Substructuring for Eigenvalue and Frequency Response Calculations joint work with X. S. Li, C. Yang, P. Husbands, E. Ng LBNL W. Gao Fudan University, China J. H. Ko Konkuk University, South Korea

EIG and FRA • EIG (eigenvalue problem) σ K σ q = λ σ Mq Kq = λMq − → where K σ = K − σM, λ σ = λ − σ • FRA (frequency response calculation) H ( ω ) = l T ( K − ω 2 M + i ωD ) − 1 b = l T [ γ 1 ( ω ) K σ + γ 2 ( ω, σ ) M ] − 1 b ↑ σ, D = αK + βM where γ 1 ( ω ) = 1 + i ωα γ 2 ( ω, σ ) = − ω 2 + σ + i ω ( σα + β ) ,

Substructuring methods • Eigenvalue and frequency response calculations are ubiquitous. • Substructuring techniques dates back to the 1960s (CMS) • Substructuring holds great promoise for solving extremely large scale problems, e.g., AMLS • Issues as the techniques extended for broader applications: – Extraction of arbitrary eigenmodes – High frequency response calculations – Accuracy enhancement – Performance tuning

Outline of Part I 1. Substructure partition and reduction (projection) 2. ASEIG : eigenvalue calculation 3. ASFRA : frequency response calculation 4. Remarks

Substructure partition I.1 • Single-level ⎡ ⎤ ⎡ ⎤ K σ K σ M 11 M 13 ⎢ 11 13 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ K σ = ⎢ K σ 22 K σ ⎥ ⎢ ⎥ M = M 22 M 23 ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ 23 ⎢ ⎥ ⎢ ⎥ ⎣ ⎦ ⎣ ⎦ K σ 31 K σ 32 K σ M 31 M 32 M 33 33 • Multi-level (nested dissection) �� 1 �� 2 �� 3 �� 4 �� 5 �� 15 �� 6 �� 7 �� 8 �� 7 14 �� 9 �� 10 �� 11 �� 6 �� 3 10 13 �� 12 �� 13 �� 14 �� 11 12 1 2 4 5 8 9 �� 15 �� e.g., by M E T I S of Karypis and Kumar

Substructure reduction (projection) I.1 • Block elimination matrix L ⎡ ⎤ ⎡ ⎤ K σ � M 11 M 13 ⎢ 11 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ K σ = L − T K σ L − 1 = M = L − T ML − 1 = ⎢ K σ ⎥ ⎢ ⎥ � � � ⎦ , M 22 M 23 ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ 22 ⎢ ⎥ ⎢ ⎥ ⎣ ⎣ ⎦ K σ � � � � M 31 M 32 M 33 33 • Extraction of local modes ⎡ ⎤ ( K σ S 1 ← − 11 , M 11 ) ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ( K σ S m = S 2 ← − 22 , M 22 ) ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎣ ⎦ K σ � � S 3 ← − ( 33 , M 33 ) • Subspace projection onto span { S m } ⎡ ⎤ ⎡ ⎤ k σ m 11 m 13 � ⎢ 11 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ K σ m = S T K σ S m = M m = S T ⎢ k σ ⎥ ⎢ ⎥ � � ⎦ , MS m = m 22 m 23 ⎢ ⎥ ⎢ � ⎥ ⎢ ⎥ ⎢ ⎥ m 22 m ⎢ ⎥ ⎢ ⎥ ⎣ ⎣ ⎦ k σ � m 31 m 32 m 33 � � � 33

Eigenvalue calculation I.2 • Projected (smaller) eigenvalue problem: K σ m Φ = M m ΦΘ σ , • “Global” eigenpairs: Θ σ = diag ( θ σ ) = diag (Θ σ l , Θ σ n , Θ σ r ) � Φ l Φ n Φ r � Φ = ( φ ) = • Retained modes: Φ n ( determined by “global cutoff values” ) � Φ l Φ r � Truncated modes: Φ t = • Ritz pairs (Θ σ n + σ, L − 1 S m Φ n ) ≈ ( λ, q ) ’s of ( K, M ) • Why does it work, ... [Yang et al ], [Voss et al ], ...

AS algorithm for eigenvalue computation I.2 0. Substructure partition, if necessary 1. Block elimination and congruence transformation 2. Mode selection/computation for sub-structures and separators 3. Subspace assembling 4. Projection calculation 5. Projected eigenvalue problem

AS implementation I.2 • Major operations: 1. Congruence transformations and projection 2. projected eigenvalue problem • Cost 1. flops: more than a single sparse Cholesky factorization 2. storage: Block Cholesky factors + projected matrices + ... But no triangular solvers, no (re)-orthogonalization ... vs. SIL • ASEIG package [ACM TOMS ’08] 1. interleaves steps 1-4 computations 2. recomputes some of intermediate matrix blocks instead of storing 50% of memory saving trade with 15% recompute time

Linear and Nonlinear SP 2 Methods for Large Scale Eigenvalue - PowerPoint PPT Presentation

Linear and Nonlinear SP 2 Methods for Large Scale Eigenvalue Calculations Zhaojun Bai University of California, Davis Workshop on Recent Advances in Numerical Methods for Eigenvalue Problems NCTS, Taiwan, January 4-8, 2008 Linear and

Random Methods for Large-Scale Linear Problems, Variational Inequalities, and Convex Optimization

Scale Linear Systems Y A O H A N G L I , P H . D . D E P A R T M E N T O F C O M P U T E R S

Nonlinear Methods Data often lies on or near a nonlinear low-dimensional curve aka manifold. 27

On iterative regularization methods for nonlinear inverse problems Antonio Leito

Numerical Methods for Solving Large Scale Eigenvalue Problems Lecture 2, February 28, 2018:

Nonlinear Classifiers II 2 Nonlinear Classifiers: Introduction Classifiers Supervised

Rational Krylov methods for linear and nonlinear eigenvalue problems Mele Giampaolo

Cascadic Multilevel Methods for Cascadic Multilevel Methods for Large-Scale Ill-Posed Problems

SC13 GPU Technology Theater AmgX: Performance Acceleration for Large-Scale Iterative Methods

Nonlinear dynamics emerging in large scale neural networks with ontogenetic and epigenetic

Spacetime domain decomposition methods for linear and nonlinear diffusion problems Michel

Survey of Fast Methods for large-scale tree estimation J E S U S S A N D O V A L Introduction

Large-scale production of graphene: Introduction Large-scale production of graphene: what is

New developments of LOBPCG for large-scale nonlinear eigenvalue problems Fei Xue University of

Linear and nonlinear methods for model reduction Diane Guignard Joint work : A. Bonito, R.

Composing Nonlinear Solvers Matthew Knepley Computational and Applied Mathematics Rice

NEPTUNE Scheduling Suspendable Tasks for Unified Stream/Batch Applications Panagiotis

Counterfighting Counterfeit Detecting and taking down fraudulent webshops at a ccTLD Thymen Wabeke

Ecological transformation When ecological knowledge and transformation in the social

A Short History of Array Computing in Python Wolf Vollprecht, PyParis 2018 TOC - Array

Simple, Fast and Deterministic Gossip and Rumor Spreading Main paper by: B. Haeupler, MIT Talk

Statistical Methods used in Reactor Neutrino Experiments Xin Qian BNL 1 Reactor Neutrinos

Nutrient loads from estuaries to the coastal ocean; the role of resolution and vegetation on

Distributed Multi-GPU Computing with Dask, CuPy and RAPIDS Peter Andreas Entschev Senior System