[petsc-dev] PETSc LU vs SuperLU

[Dave Nystrom]

So, to do this, would I want to rebuild PETSc using a vendor compiler like
Intel or PGI?  Are there other performance benefits to building PETSc with a
vendor compiler such as Intel or PGI besides the benefit of linking with
vendor blas?

Xiaoye S. Li writes:
 > For SuperLU, MUMPS, UMFPACK etc, you need to link with a good BLAS, like
 > MKL you mentioned.  The internal algorithms exploit dense matrix
 > sub-blocks, and good BLAS usually make a difference.   For very sparse
 > problems, the above packages may not help, since there is not much blocking
 > to exploit, and the overhead of trying to find the blocks cannot be easily
 > offset.
 > 
 > Sherry
 > 
 > 
 > On Tue, Dec 20, 2011 at 7:28 AM, Dave Nystrom <dnystrom1 at comcast.net> wrote:
 > 
 > > I have been comparing sequential SuperLU on one of my linear solves versus
 > > PETSc LU.  I am finding SuperLU to be a little over 2x slower than PETSc
 > > LU.
 > > I was wondering if this is due to SuperLU not being tuned to my problem or
 > > if
 > > the PETSc LU algorithm performance is expected to be superior to that of
 > > SuperLU in general.  I did play around with the reordering options for
 > > SuperLU but did not find anything superior to the defaults.  I was also
 > > wondering if building PETSc and its external packages with another compiler
 > > such as PGI or Intel might result in higher performance in this regard.  Or
 > > whether using a vendor blas like MKL would speed up SuperLU.  Or perhaps
 > > the
 > > interface of SuperLU to PETSc results in some extra data copying that is
 > > the
 > > difference.
 > >
 > > Does anyone have any idea why SuperLU might be that much slower than PETSc
 > > LU?
 > >
 > > I also tried spooles and that was just a little slower than PETSc LU.  And
 > > I
 > > tried MUMPS and that seg faulted after my problem had been running over an
 > > hour.  This particular problem was running for less than 3 minutes with
 > > PETSc
 > > LU.
 > >
 > > I would be interested in any suggestions of things to try to speed up my LU
 > > solve with either PETSc or any of the external packages.  Right now, I'm
 > > just
 > > doing serial, single node calculations.
 > >
 > > Thanks,
 > >
 > > Dave
 > >