On Fri, Dec 9, 2011 at 6:13 PM, Xiangdong Liang <xdliang at gmail.com> wrote:
> On Fri, Dec 9, 2011 at 7:02 PM, Barry Smith <bsmith at mcs.anl.gov> wrote: > > > > On Dec 9, 2011, at 5:55 PM, Xiangdong Liang wrote: > > > >> Hello everyone, > >> > >> I am solving Ax=b with sparse direct solver Pastix. I have two > >> equivalent A's (upto these zero entries): A1 and A2. A1 is generated > >> with ignor_zero_entries and A2 is without this option. For example A1 > >> has 9 millions nonzeros, while A2 has 10 millions zeros. When I solve > >> them with Pastix, I found the time for solving sparser A1 actually is > >> longer (10%--20% worse) than A2. Does anyone have thoughts on this? > > > > Completely possible and not particularly surprising. The amount of > work required for sparse LU depends in an incrediably complicated way on > the nonzero structure of the matrix, it is only very very minorly related > to the number of nonzeros in the matrix. > > One fun thing to check would be the number of nonzeros in the factor > of A1 and the number of nonzeros in the factor of A2. (I'm not sure if > Pastix has a way to check this). > > Should the nnz in the factor of A1 be the same as nnz in the factor of > A2 since A1 and A2 are the same matrices except some zeros (due to > ignore_zero_entries in the matrix assembling)? > nnz is structural, and has nothing to do with numeric values. Matt > Xiangdong > > > > > Barry > > > >> Thanks. > >> > >> Xiangdong > >> > >> P.S. The time I count is only for Spare LU solving (not including the > >> matrix assembling time). > > > -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20111209/94dd4e02/attachment-0001.htm>
