Domain decomposition schemes with high-order accuracy and unconditional stability

Similar documents
Numerical Solutions to Partial Differential Equations

A FINITE DIFFERENCE DOMAIN DECOMPOSITION ALGORITHM FOR NUMERICAL SOLUTION OF THE HEAT EQUATION

Introduction to numerical schemes

Block-Structured Adaptive Mesh Refinement

Finite Difference Methods for Boundary Value Problems

Finite difference method for elliptic problems: I

Chapter 5 HIGH ACCURACY CUBIC SPLINE APPROXIMATION FOR TWO DIMENSIONAL QUASI-LINEAR ELLIPTIC BOUNDARY VALUE PROBLEMS

Numerical Solutions to Partial Differential Equations

A stencil of the finite-difference method for the 2D convection diffusion equation and its new iterative scheme

Finite Differences for Differential Equations 28 PART II. Finite Difference Methods for Differential Equations

SECOND-ORDER FULLY DISCRETIZED PROJECTION METHOD FOR INCOMPRESSIBLE NAVIER-STOKES EQUATIONS

Parallel Galerkin Domain Decomposition Procedures for Parabolic Equation on General Domain

High-order ADI schemes for convection-diffusion equations with mixed derivative terms

Finite difference method for heat equation

An Accurate Fourier-Spectral Solver for Variable Coefficient Elliptic Equations

Tutorial 2. Introduction to numerical schemes

HIGH-ORDER ACCURATE METHODS BASED ON DIFFERENCE POTENTIALS FOR 2D PARABOLIC INTERFACE MODELS

A Simple Compact Fourth-Order Poisson Solver on Polar Geometry

HIGH-ORDER ACCURATE METHODS BASED ON DIFFERENCE POTENTIALS FOR 2D PARABOLIC INTERFACE MODELS

Optimal Interface Conditions for an Arbitrary Decomposition into Subdomains

ECE539 - Advanced Theory of Semiconductors and Semiconductor Devices. Numerical Methods and Simulation / Umberto Ravaioli

Block-tridiagonal matrices

A CCD-ADI method for unsteady convection-diffusion equations

Convergence Behavior of a Two-Level Optimized Schwarz Preconditioner

An Efficient Algorithm Based on Quadratic Spline Collocation and Finite Difference Methods for Parabolic Partial Differential Equations.

An additive average Schwarz method for the plate bending problem

Code: 101MAT4 101MT4B. Today s topics Finite-difference method in 2D. Poisson equation Wave equation

Two-parameter regularization method for determining the heat source

Numerical Solutions to Partial Differential Equations

Lecture 4.2 Finite Difference Approximation

Splitting methods with boundary corrections

ON APPROXIMATION OF LAPLACIAN EIGENPROBLEM OVER A REGULAR HEXAGON WITH ZERO BOUNDARY CONDITIONS 1) 1. Introduction

A fourth-order finite difference scheme for the numerical solution of 1D linear hyperbolic equation

FDM for parabolic equations

Efficient smoothers for all-at-once multigrid methods for Poisson and Stokes control problems

Domain decomposition for the Jacobi-Davidson method: practical strategies

A FAST SOLVER FOR ELLIPTIC EQUATIONS WITH HARMONIC COEFFICIENT APPROXIMATIONS

Synchronization and Bifurcation Analysis in Coupled Networks of Discrete-Time Systems

= ( 1 P + S V P) 1. Speedup T S /T V p

Chapter Two: Numerical Methods for Elliptic PDEs. 1 Finite Difference Methods for Elliptic PDEs

A SEMI-LAGRANGIAN RUNGE-KUTTA METHOD FOR TIME-DEPENDENT PARTIAL DIFFERENTIAL EQUATIONS

Convergence and Error Bound Analysis for the Space-Time CESE Method

arxiv: v1 [math.na] 29 Feb 2016

PIECEWISE LINEAR FINITE ELEMENT METHODS ARE NOT LOCALIZED

Introduction to the Numerical Solution of IVP for ODE

PARTITION OF UNITY FOR THE STOKES PROBLEM ON NONMATCHING GRIDS

Random Graphs. EECS 126 (UC Berkeley) Spring 2019

Multi-Factor Finite Differences

Finite Difference Method of Fractional Parabolic Partial Differential Equations with Variable Coefficients

PARAREAL TIME DISCRETIZATION FOR PARABOLIC CONTROL PROBLEM

Natural Boundary Integral Method and Its Applications

INTRODUCTION TO FINITE ELEMENT METHODS

2 Two-Point Boundary Value Problems

Numerical Solutions to Partial Differential Equations

arxiv: v1 [math.na] 11 Jul 2018

A space-time Trefftz method for the second order wave equation

EXISTENCE OF SOLUTIONS TO THE CAHN-HILLIARD/ALLEN-CAHN EQUATION WITH DEGENERATE MOBILITY

A FOURTH-ORDER ACCURATE FINITE DIFFERENCE SCHEME FOR THE EXTENDED-FISHER-KOLMOGOROV EQUATION. Tlili Kadri and Khaled Omrani

Finite Difference Method

Krylov Implicit Integration Factor Methods for Semilinear Fourth-Order Equations

An Exponential High-Order Compact ADI Method for 3D Unsteady Convection Diffusion Problems

Construction of a New Domain Decomposition Method for the Stokes Equations

ON SPECTRAL METHODS FOR VOLTERRA INTEGRAL EQUATIONS AND THE CONVERGENCE ANALYSIS * 1. Introduction

Non-Conforming Finite Element Methods for Nonmatching Grids in Three Dimensions

A local Crank-Nicolson method for solving the heat equation. Abdurishit ABUDUWALI, Michio SAKAKIHARA and Hiroshi NIKI

A Domain Decomposition Based Jacobi-Davidson Algorithm for Quantum Dot Simulation

A Fast High Order Algorithm for 3D Helmholtz Equation with Dirichlet Boundary

Numerical Algorithms for Visual Computing II 2010/11 Example Solutions for Assignment 6

Numerical methods for weak solution of wave equation with van der Pol type boundary conditions

CONVERGENCE OF GAUGE METHOD FOR INCOMPRESSIBLE FLOW CHENG WANG AND JIAN-GUO LIU

A parallel algorithm for the heat equation with derivative boundary conditions

On an Approximation Result for Piecewise Polynomial Functions. O. Karakashian

Rational Chebyshev pseudospectral method for long-short wave equations

Preconditioning for Nonsymmetry and Time-dependence

Additive Schwarz Methods for Hyperbolic Equations

The Milne error estimator for stiff problems

Numerical Solutions to Partial Differential Equations

Compact High Order Finite Difference Stencils for Elliptic Variable Coefficient and Interface Problems

Exponentially Fitted Error Correction Methods for Solving Initial Value Problems

Finite Difference Method for PDE. Y V S S Sanyasiraju Professor, Department of Mathematics IIT Madras, Chennai 36

An interior-point trust-region polynomial algorithm for convex programming

ASYMPTOTICALLY EXACT A POSTERIORI ESTIMATORS FOR THE POINTWISE GRADIENT ERROR ON EACH ELEMENT IN IRREGULAR MESHES. PART II: THE PIECEWISE LINEAR CASE

Ultraconvergence of ZZ Patch Recovery at Mesh Symmetry Points

Boundary Value Problems and Approximate Solutions

A Hybrid Method for the Wave Equation. beilina

THE METHOD OF LINES FOR PARABOLIC PARTIAL INTEGRO-DIFFERENTIAL EQUATIONS

Solving a class of nonlinear two-dimensional Volterra integral equations by using two-dimensional triangular orthogonal functions.

Performance Analysis of Parallel Alternating Directions Algorithm for Time Dependent Problems

Error formulas for divided difference expansions and numerical differentiation

On the efficiency of the Peaceman-Rachford ADI-dG method for wave-type methods

Analysis of Two-Grid Methods for Nonlinear Parabolic Equations by Expanded Mixed Finite Element Methods

CRANK-NICOLSON FINITE DIFFERENCE METHOD FOR SOLVING TIME-FRACTIONAL DIFFUSION EQUATION

Local Mesh Refinement with the PCD Method

Draft TAYLOR SERIES METHOD FOR SYSTEM OF PARTICLES INTERACTING VIA LENNARD-JONES POTENTIAL. Nikolai Shegunov, Ivan Hristov

arxiv: v1 [math.na] 15 Nov 2017

MATH 425, HOMEWORK 5, SOLUTIONS

AMS subject classifications. Primary, 65N15, 65N30, 76D07; Secondary, 35B45, 35J50

DYNAMICS IN 3-SPECIES PREDATOR-PREY MODELS WITH TIME DELAYS. Wei Feng

Hamburger Beiträge zur Angewandten Mathematik

DOMAIN DECOMPOSITION APPROACHES FOR MESH GENERATION VIA THE EQUIDISTRIBUTION PRINCIPLE

Transcription:

Domain decomposition schemes with high-order accuracy and unconditional stability Wenrui Hao Shaohong Zhu March 7, 0 Abstract Parallel finite difference schemes with high-order accuracy and unconditional stability for solving parabolic equations are presented. The schemes are based on domain decomposition method, i.e., interface values between subdomains are computed by the explicit scheme; interior values are computed by the implicit scheme. The numerical stability and error are derived in the H norm in one dimensional case. Numerical results of both one and two dimensions examining the stability, accuracy, and parallelism of the procedure are also presented. Keywords: Domain decomposition, Finite difference, Parabolic equation, High-order accuracy, Unconditional stability. Introduction Domain decomposition is a powerful tool for devising parallel methods to solve time-dependent partial differential equations. There is rich literature on domain decomposition finite difference methods for solving parabolic equations on parallel computers. For the non-overlapping domain decomposition methods, the explicit nature of the calculation at the interface of sub-domain leads some domain decomposition schemes to be conditionally stable, which implies that they have to suffer from temporal step-size restrictions (see []-[5]). Schemes with unconditional stability as well as high-order accuracy being desired in the applications, many investigators have turned to improve the stability of the domain decomposition method. For example, the corrected explicit-implicit domain decomposition algorithms were presented in [6] and [7]. By adding the correction step to explicit-implicit domain decomposition methods, updating the interface solutions at each time level, the corrected methods were proved to Department of Applied and Computational Mathematics and Statistics, University of Notre Dame, Notre Dame, IN 46556 USA (whao@nd.edu). School of Mathematics Science and LPMC, Nanai University, Tianin, 30007, China (shhzhu@nanai.edu.cn).

have the unconditional stability. While the needless corrected domain decomposition schemes with unconditional stability were presented in [8] and [9]. All of these methods with unconditional stability reach the second order accuracy at most. The purpose of this paper is to present the domain decomposition finite difference procedure with third-order accuracy and unconditional stability. We first consider the following Dirichlet boundary problem U t U = 0, x (0, ), t (0, T], x (.) U(0, t) = U(, t) = 0, t (0, T], (.) U(x, 0) = U 0 (x), 0 x, (.3) where the initial function U 0 (x) satisfies the boundary condition, i.e., U 0 (0) = U 0 () = 0. Then we extend the method to the problem of two dimensional space. We will introduce two new finite difference schemes for solving (.)-(.3) in Section, and setch the domain decomposition procedure, the numerical stability and convergence in Section 3. In section 4, the proof of the unconditional stability and the error estimate will be given. Numerical examples and examination of the algorithm will be provided in Section 5. In Sections 6 and 7, we extend the method to the problem of two dimensional space and test some examples. Two new finite difference schemes Taing the usual h, τ mesh in x and t, and denoting the approximate value of U(x, t n ) U n by u n, where x = h and t n = nτ, we define the following operators + u n = un + un, u n = un un, τu n = τ (un un ). It is well nown that the following Taylor expansion resulting in the fully implicit finite difference scheme is valid. τ U + U ( ) τ h = + h U t (x, t ) + O(τ + h 4 ). (.4) Noticing that U t (x, t ) = U U n + Un τ + O(τ), (.5) and substituting U t (x, t ) into (.4), we obtain τ U + U ( ) τ h + + h U U n + Un τ = O(τ + h 4 + τh ). (.6)

3 Replacing by + and in (.5), we have and U+ = Un + Un + + τ U t (x +, t ) + O(τ 3 ) U = Un Un + τ U t (x, t ) + O(τ 3 ). Substituting U± into (.6), we can get τ U Un + Un + U + U n Un h + ( = τ U h t (x +, t ) + U t (x, t ) ( ) τ + h U ) + O(τ + h 4 + τh + τ3 U n + Un τ h ). (.7) By omitting the high order term, (.6) and (.7) yield two new finite difference schemes for (.) : τ u + u ( τ h + + h ( τ + h τ u + u h + ) u ) u u n + un τ = 0, (.8) u n + un τ r( τ u n + τ u + ) r( τu n τ u ) = 0, (.9) where r = τ. From the derivation of (.8) and (.9), we now that the truncation errors of (.8) and (.9) are O(τ +h 4 +τh h ( ) τ ) and O h +τ +h 4 +τh + τ3 h respectively. If r is any positive real number, the truncation errors become O(h 4 ) and O(h ). 3 Domain decomposition procedure and main results Suppose Jh =, Nτ = T. For simplicity, we will consider a domain decomposition which involves in decomposing (0, ) into only two subdomains, (0, x) and (x, ), where x = x for some integer ( < < J ). We use the explicit scheme (.9) to compute the solution value u and the implicit scheme (.8) to compute other solution values u ( ) respectively. The system can be written as L(u ) = 0, n N, 0 < < J, u = 0, = 0, J, (3.0)

4 where the linear operator L is L(u ) = τ u + u ( ) τ h + + h u u n + un τ,, τ u + u ( ) τ h + + h u u n + un τ r( τ u n + τ u + ) r( τu n τ u ), =. (3.). The resulting system of equations decouples into two disoint sets of equations corresponding to the subdomains. These systems can be solved in parallel. We are now in position to state two main theorems of this paper, which will be proved in the forthcoming sections. For the discrete function u n = {u n = 0,,..., J, u n 0 = un J = 0}, define u n = u n h, + u n = + u n h. = We have the following theorems. Theorem 3. (Stability) For any given r > 0, the finite difference solutions of the parallel scheme (3.0)-(3.) satisfy max +u n n =0 r + r + ( +u + u u 0 ). Theorem 3. (Convergence) Let e n = U(x, t n ) u n. For any given r > 0, the finite difference solutions of the parallel scheme (3.0)-(3.) satisfy max n +e n C( + e + e e 0 + h 4 ), where C is a positive constant independent of h and τ. Since the finite difference scheme (3.0) has three time levels, besides taing u 0 = U 0(h), thus e 0 = 0, we need to find other methods to solve u,. In order to match high order accuracy, we can use either fourth order explicit schemes such as the impact scheme or the high-order parallel iterative method [0] to compute u, i.e., let e O(h4 ). Then according to Theorem 3., scheme (3.0) will reach third-order accuracy. 4 Proof of stability and convergence We first state three auxiliary lemmas. The stability and convergence results are then derived.

5 Lemma 4. (Discrete Poincare Inequality) For the discrete function u n = {u n = 0,,..., J, un 0 = u n J = 0}, there exists u n + u n h. Lemma 4. (Discrete Green Theorem) If u and v are discrete functions on the set {x = 0,,, J}, then we have J u + v = v u u 0 v + u J v J. = = Lemma 4.3 For any given r > 0, f L (0, T; L (0, )), the finite difference solutions of system L(u ) = f with the Dirichlet boundary condition satisfy ( max +u n + u + n r + ) u r + u 0 + Th max r n fn. Lemma 4. and Lemma 4. are proved in []. Before proving the stability and convergence, we will give the proof of Lemma 4.3. 4. Proof of Lemma 4.3 Denoting w = u boundary condition as w r + u u n, we can rewrite L(u ) = f with the Dirichlet ( + w r + u + + ) ( r + r (w ) (w w n ) r(w+ n w + ) r(wn w ) = f u = 0, = 0, J. w n ) = f τ,, τ, =, Multiplying the above equations by w h, =,,, J and summing them up respect to, we have = (w rw ) h r = ( w + u h + + ) r (w+ n w + )h rw (w n w From Lemma 4., we have = )h = = (w w n )w h f w τh. (4.)

6 J w + u = u w = = = ( + u ) + ( + u n ) ( + u + u n ). =0 Then (4.) becomes w + r [ =0 ( + + )[ r ( + u = (w +rhw (w + + w ) = rhw =0 (w+ n + wn ) + ) h =0 =0 ] ( + u n ) h + r ) h (w n ) h + = w f = ( (w ) + (w+ n rh ) + (w τh ) + (w n ) = ) + = =0 ( + u (w ] w n ) h (w ) + (f + u n ) h τ) rh(w ) + rh (wn + ) + (w n ) + w + τ f. (4.3) Noting that = r r ( + u + u n ) h + rhw (w + + w =0 =0,, ( + w we can simplify (4.3) as follows w + r ( + 4 + 4r [ (w + +rh ) + (w ) [ (w ) + h + rh ) + (w [ + u + u n ] + r ) ) ( w w n + w w n ) ) =0,, ] + rh(w ), (wn + ) + (w n ] ) τ f ( + w ) h h

7 Since w 0, we have =0,, ] ( r [ + u + u n + [ (w + +rh ) + (w ) Summing up respect to n, we get Thus ( r +u + 4 + r +u + ( 4 + 4r + u + u + Lemma 4.3 is proved. 4. Proof of stability ( + w ) h 0, and w w n 0, then 4 + 4r (wn + ) + (w n ) ) ( w w n ) ] τ f ) w + rh (w + ) + (w ) 4r ) w + rh (w + ) + (w ) + Tτ max n fn. ( r + ) u r + u 0 + Th max r n fn. Let f = 0 in Lemma 4.3, thus Theorem 3. is proved. 4.3 Proof of convergence According to the parallel scheme (3.0)-(3.), the errors e ( n < N) follows the below equations: τ e + e ( τ h + + h τ e + e ( τ h + + h ) e ) e e n + en τ = G,, e n + en τ r( τ e n + τe + ) r( τe n τe ) = Φ h + G, =, e = 0, = 0, J, where Φ = r U t (x, t ), G O(h 4 ). In order to get the error estimates for e, we assume that e = p + q, where p and q are the solutions of the following problems respectively.

8 problem I τ p + p ( τ h + + h τ p + p ( τ h + + h ) p ) p p n + pn τ = G,, p n + pn τ r( τ p n + τp + ) r( τp n τp ) = G, =, (4.4) p = 0, = 0, J, p 0 = e0, p = e,. problem II τ q + q ( τ h + + h τ q + q ( τ h + + h ) q ) q q n + qn τ = 0,, q n + qn τ r( τ q n + τ q + ) r( τq n τ q ) = Φ h, =, q = 0, = 0, J, q 0 = q = 0,. From lemma 4.3, we can obtain the estimate for p, i.e., ( + p n + p + r + ) p r + p 0 + Th max r n Gn ( r + ) r + ( + e + e e 0 ) + O(h 0 ). (4.5) For estimating q, we first consider q ( n N )which satisfies the following equations Then the formula of q q = + q h = 0,, + q h = Φ h, q 0 = q J = 0. is 0, = 0, J, J Φ h 4,, J + J J Φ h 4, < < J. i=

9 Hence + q O(h 4 ) and q q n O(τh 3 ). Define q = q q, then we have τ q + q ( τ h + + h τ q + q ( τ h + + h ) q ) q q n + qn τ = R,, q n + qn τ r( τ q n + τ q + ) r( τ q n τ q ) = R, =, q = 0, = 0, J, q 0 = q 0, q = q,, where R = ( τ τ q + + h ( τ τ q + + h r( τ q n τ q ) q ) q ), =, q n + qn τ,, q n + q n τ r( τ q n + τ q + ) and R O(h 3 ). From Lemma 4.3, we can obtain the estimate of + q. ( + q + q + r + ) q r + q 0 + Th max r n Rn O(h 8 ). Thus, Combining with (4.5), we get + q + q + + q O(h 4 ). + e + p + + q C( + e + e e 0 + h 4 ), where C is a positive constant independent of h and τ. Theorem 3. is proved. 5 Numerical experiments In this section, some numerical results are presented to show the stability, accuracy, and parallelism of the scheme described above, and the computational costs are also presented. All the experiments are run on a cluster consisting of a manager that uses one core of a Xeon 540 processor and up to 3 computing nodes, each containing two Xeon 540 processors running 64-bit Linux, i.e., each node consists of 8 processing cores.

0 We consider the problem defined in equations (.)-(.3) with U 0 (x) = sin(πx). Obviously the exact solution of the equations is U(x, t) = e πt sin(πx). First, we verify the stability of the scheme by taing the step size h = 0 3, r =, 0, 00, 000 with two subdomains. FIG. clearly shows that the norm of u n doesn t occur blowing up even if r is large enough. This explains the unconditional stability of the scheme. r= r=0 u n 0.5 u n 0.5 0 0 0.5 time t r=00 0 0 0.5 time t r=000 u n 0.5 u n 0.5 0 0 0.5 time t 0 0 0.5 time t Figure : The infinity norm of u n v.s. t Second, we examine the numerical errors in the solutions. Table shows that the errors for each case are roughly of the same order of magnitude, and the errors appear to be O(h 3 ) in each case. Third, we test the speed-up for the scheme. Here we tae h = 0 5, τ = 0 6, i.e., r = 0 4, T = 0.5 and list the time for computing and the speed-up in Table 5, which shows that the scheme has a coemptive parallelism. 6 Extension to two dimensional case In this section, U(x, y, t) will be a solution of the following Dirichlet boundary problem on Ω = (0, ) (0, ),

Table : Numerical errors for different grid points r = 0, T = processors 4 processors 8 processors J u n U n /h 3 J u n U n /h 3 J u n U n /h 3 000.068 000 4.05448 000 8.07385 000.0675 000 4.05336 000 8.0776 4000.0636 4000 4.04938 4000 8.0688 8000.9499 8000 3.93590 8000 7.9563 r = 00, T = processors 4 processors 8 processors J u n U n /h 3 J u n U n /h 3 J u n U n /h 3 000.06703e+0 000 4.04766e+0 000 8.0437e+0 000.0679e+0 000 4.0599e+0 000 8.06955e+0 4000.06690e+0 4000 4.05307e+0 4000 8.076e+0 8000.0680e+0 8000 4.05434e+0 8000 8.07356e+0 r = 000, T = processors 4 processors 8 processors J u n U n /h 3 J u n U n /h 3 J u n U n /h 3 000.0883e+04 000 4.0444e+04 000 8.03459e+04 000.03740e+04 000 3.9388e+04 000 8.03843e+04 4000.06360e+04 4000 4.03957e+04 4000 8.0579e+04 8000.06634e+04 8000 4.0537e+04 8000 8.06558e+04 U U = 0, (x, y) Ω, t (0, T], t (6.6) U(x, y, t) = 0, (x, y) Ω, t (0, T], (6.7) U(x, y, 0) = U 0 (x, y), (x, y) Ω, (6.8) where the initial function U 0 (x, y) satisfies the boundary condition, and U = U x + U. We extend the domain decomposition method stated in Section 3 y to the above problem. Tae Ω = {(x, y) Ω : x < x}, Ω = {(x, y) Ω : x > x}. Let x i = ih same as in Section 3, and let y = h. Suppose that there exists an integer such that x = x (see Fig. ). In analogy with Section 3 we call points (x i, y, t n ) as boundary points if (x i, y ) Ω, and interface points if i =. Otherwise, we call them interior points. The values u n i, will approximate U(x i, y, t n ) Ui, n. We denote + u i, = u i+, u i, + u i,, δ + δ u i, = u i,+ u i, + u i,,

Table : Comparison of time for different processors Processors n Time t n (seconds) Speed-up t /t n 486.66-450.8.98 4 50.9 3.89 8 6.74 7.8 6 35.05 4.96 3 65.85 9.3 64 00.45 48.40 Figure : domain decomposition for two dimensional case and Then Thus we have τ u n i, = un i, un i,. τ + U i, h = U x (x i, y, t) + h 4 U x 4 (x i, y, t) + O(h 4 ), δ + δ U i, h = U y (x i, y, t) + h 4 U y 4 (x i, y, t) + O(h 4 ), δ + δ + U i, h 4 = 4 U x y (x i, y, t) + O(h ). + U i, + δ + δ U i, h + 6h δ +δ + U i, = U(x i, y, t) + h ( 4 U x 4 (x i, y, t) + 4 U x y (x i, y, t) + 4 U y 4 (x i, y, t) ) + O(h 4 ) = U(x i, y, t) + h U t (x i, y, t) + O(h 4 ). (6.9)

3 Moreover, τ U i, = U t (x i, y, t ) τ U t (x i, y, t ) + O(τ ). (6.0) Subtract (6.0) with (6.9), we obtain τ Ui, + Ui, + δ + δ Ui, h 6h δ +δ + Ui, ( ) τ = + h U t (x i, y, t ) + O(τ + h 4 ). Therefore, the fourth order finite difference scheme for (6.6) is τ u i, + u i, + δ + δ u i, h 6h δ +δ + u i, ( ) τ + + h u i, u n i, + un i, τ = 0, (6.) which is a nine-point finite difference scheme with three levels (see Fig. 3). Figure 3: nine point scheme Denote u i, = u i, + u i,+ + u i+, + u i+,+ 4u i, = + u i, + δ + δ u i, + δ +δ + u i,. Then (6.) is equivalent to τ u i, u i, + + u i, + δ + δ u i, 3h + ( τ + h ) u i, u n i, + un i, τ = 0,

4 which can be used for computing u i, when i. When i =, u i±, in (6.) are replaced by u n i±, un i±,. In this case, we need to solve a tridiagonal matrix using Thomas algorithm for computing u,. It is straightforward to extend the two-subdomain results to many subdomains on x direction by cutting the whole domain into vertical strips. Moreover, we can also extend the domain decomposition scheme to three dimensional case by dividing x axis into many subdomains. 7 Two dimensional numerical experiments We consider the problem defined in equations (6.6)-(6.8) with U 0 (x, y) = sin(πx)sin(πy). Obviously the exact solution of the equations is U(x, y, t) = e πt sin(πx)sin(πy). The computing time and the speed-up of the scheme (6.9-6.0) are shown in Table 3 with h = 0 3, τ = 0 and T =. Table 3: Comparison of time for two dimensional case Processors n Time t n (seconds) Speed-up t /t n 0h3m9s - 6h0m9s.7 4 h5m4s 3.6 8 h3m54s 7.4 6 4m47s 4.53 3 m3s 9.06 64 m58s 47.9 7. Lota-Volterra system The competitive Lota-Volterra equations are a simple model of the population dynamics of species competing for some common resource. Given two populations, u and v, with logistic dynamics, the Lota-Volterra formulation adds an additional term to account for the species interactions. Thus the competitive Lota-Volterra equations are: u t = u + u( v), (x, y) Ω v t = v v( u), (x, y) Ω v = u = 0, (x, y) Ω. We choose random values for u(x, y) and v(x, y) as initial conditions shown in FIG 4. We tae h = 0.00, τ = 0.0, T = and show the computing time in Table 4. FIG 5 shows the numerical solution for T =.

5 u(x,y,t) time t=0 0.8 0.6 0.8 0.6 y 0.4 0.4 0. 0. 0. 0. 0.3 0.4 0.5 0.6 0.7 0.8 0.9 x v(x,y,t) time t=0 0.8 0.6 0.8 0.6 y 0.4 0.4 0. 0. 0. 0. 0.3 0.4 0.5 0.6 0.7 0.8 0.9 x Figure 4: Initial condition for the Lota-Volterra system Table 4: Comparison of time for Lota-Volterra system Processors n Time t n (seconds) Speed-up t /t n h5m3s - h45m8s.86 4 6hm36s 3.5 8 h58m57s 7.37 6 h33m37s 4.6 3 45m8s 8.85 64 8ms 46.5 8 4m33s 90. 56 7m8s 75.6 Conclusion We presented domain decomposition finite difference schemes with unconditional stability and third-order accuracy for the parabolic system. Error estimate and stability of the numerical solutions have been derived for one dimensional case. The scheme is easy to implement the parallelism and is extended in

6 y 0.8 0.6 0.4 0. u(x,y,t) time t= x 0 5 3.5 3.5.5 0.5 0.8 0. 0. 0.3 0.4 0.5 0.6 0.7 0.8 0.9 x v(x,y,t) time t= x 0 6 0 y 0.6 0.4 0. 8 6 4 0. 0. 0.3 0.4 0.5 0.6 0.7 0.8 0.9 x Figure 5: Numerical solutions for T = two and three dimensional case. The numerical results demonstrate the good performance of the parallel scheme, namely, unconditional stability, the third order accuracy and high degree of parallelism. References [] C.N. Dawson, Q. Du, and T. F. Dupont, A finite difference domain decomposition algorithm for numerical solution of the Heat equation,math. Comp., Vol. 57, 63 7 (99). [] G. Yuan, S. Zhu and L. Shen, Domain decomposition algorithm based on the group explicit formula for the heat equation, Int. J. Comput. Math, Vol. 8, 95 306 (005). [3] S. Zhu, Z. Yu and J. Zhao, A high-order parallel finite difference algorithm, Applied Math. and Comp., Vol. 83, 365-37 (006). [4] C.N. Dawson and T.F. Dupont, Explicit/implicit conservative domain decomposition procedures for parabolic problems based on bloc-centered finite differences, SIAM J. Numer. Anal., Vol. 3, 045 06 (994).

7 [5] G. Yuan and L. Shen, Stability and convergence of the explicit-implicit conservative domain decomposition procedure for parabolic problems, Computers and Math. with Applications, Vol. 47, 793-80 (004). [6] H. Shi and H. Liao, Unconditional stability of corrected explicit-implicit domain decomposition algorithms for parallel approximation of heat equations, SIAM J. Numer. Anal., Vol. 44, 584-6 (006). [7] H. Liao, H. Shi and Z. Sun, Corrected explicit-implicit domain decomposition algorithms for two-dimensional semilinear parabolic equations, Science in China Series A, Vol. 5, 36-388 (009). [8] Z. Sheng, G. Yuan and X. Hang, Unconditional stability of parallel difference schemes with second order accuracy for parabolic equation, Applied Math. and Comp., Vol. 84, 05-03 (007). [9] S. Zhu, Conservative domain decomposition procedure with unconditional stability and second-order accuracy, Applied Math. and Comp., Vol.6, 375-38 (00). [0] W. Hao and S. Zhu, Parallel iterative methods for parabolic equations, Int. J. Comput. Math, Vol. 86, 43 440 (009). [] Y. Zhou, Applications of discrete functional analysis to the finite difference method, International Academic publishers (99).