Good to know. I did actually manage to get them to run after. I had to convert the mpibaij matrix to mpiaij and then to one of those formats. They had almost identically the same performance on a modern intel processor.
Thanks for the reply Gaetan On Thu, Aug 30, 2012 at 7:21 PM, Barry Smith <bsmith at mcs.anl.gov> wrote: > > You have no reason to use those formats. They are specialized for the > Cray X1 vector machine which no longer exists. They will not be faster on > another machine. Sorry for the confusion. > > Barry > > On Aug 30, 2012, at 3:51 PM, Gaetan Kenway <kenway at utias.utoronto.ca> > wrote: > > > Hi > > > > I was wondering if anyone had any experience with using these new matrix > formats in PETSc. I have a block aij matrix (with block size 5) and tried > converting to either of these types and it just trashed memory for 10 > minutes and I then killed it. Matrix assemble is 13 seconds for reference. > I then tried to construct the matrix directly with "MatCreateMPIAIJPERM" > or "MatCreateMPIAIJCRL". Neither of these functions are wrapped in > Fortran. Also the documentation for "MatCreateMPIAIJCRL" is extremely > confusing as it says that comm is set to PETSC_COMM_SELF which defeats the > purpose of a MPI matrix. > > > > Therefor, two questions: > > 1. Have any users seen a performance benefit of either of these formats > (in parallel)? > > 2. How does one actually efficiently create such matrices? (from Fortran) > > > > On a side note, is it possible somehow to indicate on the manual pages > all the functions that are not actually wrapped in Fortran? It would make > it easier to know which one you can and can't use instead of having the > compiler tell you it can't find a symbol. > > > > Thanks, > > Gaetan > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120830/6fe2b44f/attachment.html>
