Matrices contributed by The Mathworks, Inc. Some matrices come from other sources. pivtol: A small matrix from the spline toolbox that illustrates the dramatic growth factor that can occur with a pivot tolerance of 0.1 (in UMFPACK v4.0). Needs a tolerance of 0.26, for off-diagonal pivoting. Pd: An unsymmetric matrix (x=P\d) with many sparse right-hand-sides. The Pd.A matrix is 8081-by-8081. Pd.b and Pd_rhs.A are the same 8081-by-12406 matrix (the right-hand-sides). It is highly reducible to block triangular form, with 7877 diagonal blocks, the largest of which is 30-by-30. Only 9 of the blocks are bigger than 1-by-1. This is a good candidate for a solver based on dmperm. Harvard500: Web connectivity matrix from Cleve Moler (Numerical Computing with MATLAB, 2004). A(i,j) is nonzero if there is a link from page j to page i. The URL of page j is in Problem.url{j} in the MATLAB form of this problem. For more details, see Cleve Moler's book at http://www.mathworks.com/moler . The matrix is 500-by-500, with a structural rank of 233 and numerical rank of 170. tomography: An earth science problem from David Yang. Contributed by Bobby Cheng. Much like a medical tomography problem. "UMFPACK 4.0 is consistently slower than MATLAB v5." In MATLAB 7.1, [L,U,P] = lu (A) takes 0.07 seconds (using GPLU), but UMFPACK v4.3 with [L,U,P,Q] = lu (A) takes 0.11 seconds. Kuu, Muu: symmetric positive definite matrices which caused problems with UMFPACK v4.0 (in MATLAB 6.5). chol, or UMFPACK v4.3 or later, work fine.