13.07.2015 Views

AN EVALUATION OF PARALLEL MULTIGRID AS A SOLVER AND A ...

AN EVALUATION OF PARALLEL MULTIGRID AS A SOLVER AND A ...

AN EVALUATION OF PARALLEL MULTIGRID AS A SOLVER AND A ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

96 C. W. OOSTERLEE <strong>AN</strong>D T. W<strong>AS</strong>HIO}r L := f L − A L u L ;f L−1 := R L−1 r L ;u L−1 := 0;u L−1 := MGF(A L−1 , f L−1 , u L−1 , ν 1 , ν 2 , ν 3 );u L := u L + P L u L−1 ;u L := smoother(A L , f L , u L , ν 2 );r L := f L − A L u L ;f L−1 := R L−1 r L ;u L−1 := 0;u L−1 := MGV(A L−1 , f L−1 , u L−1 , ν 1 , ν 2 , ν 3 );u L := u L + P L u L−1 ;u L := smoother(A L , f L , u L , ν 2 );All components in the code have been explained in this section.2.2.2. Multigrid based on the block LU decomposition. The third algorithm,MG3, evaluated as a solver and a preconditioner, was recently proposed byReusken ([12]). It originates from several publications in which the fine grid matrixA is decomposed ([10], [11]). In matrix A the fine grid points (excluding the coarsegrid points) are ordered in the first part, the coarse grid points in the second part.The coefficient matrix on the fine grid is now described as( )A11 A(39)A =12.A 21 A 22An approximation M of the above matrix based on the block LU decomposition ismade as follows:() ( )IM =1 0 A11 A(40)12−1.A 21 A 11 I 2 0 SIf S equals the Schur complement S A (= A 22 − A 21 A −1 11 A 12 ), M is equal to A.Multigrid methods based on the formulation and splitting from (40) lead to interestingproperties of coarse grid matrices. If the fine grid matrix is an M-matrix ([20]),S will also be an M-matrix. In [12], S is an approximate Schur complement constructedfrom the Schur complement of a modified matrix A. With the modificationbased on linear interpolation, A 11 and A 12 reduce to A 11 and A 12 , where A 11 isdiagonal:(41)( )A11 AA =12; AA 21 A 11 = diag(A 11 ), A 12 = A 12 + offdiag(A 11 )P 12 .22In (41), P 12 is an operator that interpolates fine grid elements bilinearly to adjacentcoarse grid points.A damping parameter ω is used to correct an incorrect scaling of S compared toS A . So(42)S = ω −1 (A 22 − A 21 A −111 A 12 ).Another choice for S based on lumping is found in [11].We construct multigrid cycles based on the Richardson iteration of the splittingof A = M + (A − M) as in (4).

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!