Fill Reduction Algorithm Using Diagonal Markowitz Scheme with Local Symmetrization Patrick Amestoy ENSEEIHT-IRIT, France Xiaoye S. Li Esmond Ng Lawrence Berkeley National Laboratory

SIAM CSE03, Feb 10-13, Contents  Motivation  Graph models for Gaussian elimination  Minimum priority metrics  Experimental results  Summary  Add: Runtime and space complexity

SIAM CSE03, Feb 10-13, Motivation -- New Sparse LU Factorization Algorithms  Inexpensive pre/post-processing  Equilibration (or scaling)  Pre-permute rows or columns of A to maximize its diagonal  Find a matching with maximum weight for bipartite graph of A  Example: MC64 [Duff/Koster ‘99]  Iterative refinement  GESP (static pivoting) [Li/Demmel ‘98, SuperLU_DIST]  Pivots are chosen from the diagonal  Allow half-precision perturbation of small diagonals  Unsymmetrized multifrontal [Amestoy/Puglisi ‘00, MA41_NEW]  Prefer diagonal pivoting, but threshold pivoting is possible  Allow unsymmetric fronts, but dependency graph is still a tree  Diagonal is (almost) good  Struct(L’)  Struct(U)

SIAM CSE03, Feb 10-13, Existing Ordering Strategies to Preserve Sparsity  Symmetric ordering algorithms on A’+A  Greedy algorithms  e.g., minimum degree, minimum deficiency, etc.  Graph partitioning  Hybrid  Problem: unsymmetric structure is not respected!

SIAM CSE03, Feb 10-13, Structural Gaussian Elimination -- Symmetric Case i jk i jk Eliminate 1 1 i j k 1 i j k 1 i j k i k j Undirected graph After a vertex is eliminated, all its neighbors become a clique The edges of the clique are the potential fills (upper bound !)

SIAM CSE03, Feb 10-13, Structural Gaussian Elimination -- Unsymmetric Case Eliminate 1 1 r1 r2 c1 c2 c3 1 r1 r2 c1 c2 c3 c1r1 r2 c2 c3 Eliminate 1 r1 r2 c1 c2 c3 1 1 Bipartite graph After a vertex is eliminated, all the row & column vertices adjacent to it become fully connected – “bi-clique” (assuming diagonal pivot) The edges of the bi-clique are the potential fills (upper bound !)

SIAM CSE03, Feb 10-13, Ordering Algorithms Revisit  Markowitz [1957] for unsymmetric matrices  At step k, pick pivot in the trailing submatrix so that:  It has minimum, and  It is bounded by a numerical threshold  Bound the size of the rank-1 update matrix  Expensive to implement because it is mixed with numerical consideration  Examples: MA48 (HSL), etc.  “Restricted” Markowitz -- only look ahead a few candidate columns (rows) with the lowest degrees [Zlatev ‘80]  Minimum degree [Tinney/Walker ‘67]  Special case of Markowitz for SPD systems  Efficient implementation, because:  Diagonal is stable as numerical pivot  Use quotient graph as a compact representation without regard of numerical values

SIAM CSE03, Feb 10-13, Simulation Result  Order(A) vs. Order(A’+A) (Markowitz vs. min degree)  Diagonal pivoting  88 unsymmetric matrices  Mean fill ratio 0.90  Mean flops ratio 0.79  54 very unsymmetric (symmetry <= 0.5)  Mean fill ratio 0.85  Mean flops ratio 0.56

SIAM CSE03, Feb 10-13, Quotient Graph – Symmetric Case  Elements -- representative nodes of the connected components in the eliminated subgraph  Variables -- uneliminated nodes Current pivot p: If variable v adjacent to e1, it will be adjacent to p  e1 can be absorbed by p  p is representative of conn. comp. {e1, e2, p} e1 e2 pxx x x. element list = {e1, e2}. variable list v

SIAM CSE03, Feb 10-13, Quotient Graph -- Unsymmetric Case Current pivot p: Difficulty: Path length may be greater than 2 ! e1 e2 p x x x v

SIAM CSE03, Feb 10-13, Quotient Graph -- “Local Symmetrization” e1 e2 p x x x v Current pivot p: Advantage: - Path length bounded by 2 ! Disadvantage: - Lose some asymmetry - More fill ss s

SIAM CSE03, Feb 10-13, Cost of Implementation  Elimination models can be implemented using standard graphs or quotient graphs, with different cost in time & space.

SIAM CSE03, Feb 10-13, Minimum Priority Metrics  Metrics are based on “approximate degree” in the sense of AMD, can be implemented efficiently  Almost the same cost using various metrics:  Based on row & column counts:  PRODUCT (a.k.a. Markowitz), SUM, MIN, MAX, etc.  Minimum fill : areas associated with the existing cliques are deducted  …...

SIAM CSE03, Feb 10-13, Preliminary Results with Local Symmetrization  Matrices: 98 unsymmetric in structure  Metrics : based on row/column counts or fill  Solvers:  MA41_NEW : unsymmetrized multifrontal  Local symmetrization ordering is ideal for this solver  SuperLU_DIST : GESP

SIAM CSE03, Feb 10-13, Compare Different Metrics  Solver: MA41_NEW  Average fill ratio using various metrics with respect to Markowitz (product of row & col counts)

SIAM CSE03, Feb 10-13, Compare with AMD(A’+A) using Min Fill -- All Unsymmetric  MA41_NEW  SuperLU_DIST

SIAM CSE03, Feb 10-13, Compare with AMD(A’+A) using Min Fill -- Very Unsymmetric  MA41_NEW  SuperLU_DIST

SIAM CSE03, Feb 10-13, Summary  First implementation based on BQG model  Features: supervariable, element absorption, mass elimination  Using approximate degree (degree upper bound)  Tried various metrics on large collection of matrices  PRODUCT, SUM, MIN-FILL, etc.  Not a single one is universally best, MIN-FILL is often better  Local symmetrization  Cheaper to implement, harder to understand behavior  Especially suitable for unsymmetrized multifrontal, also benefit GESP  Respectable gain for very unsymmetric matrices

SIAM CSE03, Feb 10-13, Summary (con’d)  Results for very unsymmetric matrices  Future work  Work underway for a fully unsymmetric version  Extend to graph partitioning strategy

SIAM CSE03, Feb 10-13, The End

SIAM CSE03, Feb 10-13, x 2 x x x 3 x 4 x 5 x x x 6 x x 7 Example A G(A) row column