MatMult is limited by memory bandwidth. Always analyze the data movement in your algorithm and compare the achieved "sustained memory bandwidth" to STREAM. MatMult will get a high fraction (often >80%) as long as the matrix ordering allows for reasonable cache reuse. Don't bother benchmarking with a bad ordering except to make the point that it's a bad thing to do.
On Fri, Nov 9, 2012 at 12:43 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote: > Run examples that are representative of a workload you are interested in. > src/snes/examples/tutorials/ex5.c is a simple problem that is easy to > understand. Use a 3D problem if that's what you're interested in. Run with > -log_summary to get detailed information. Read the user's manual section on > profiling for more details. > > > On Fri, Nov 9, 2012 at 11:20 AM, Reuben D. Budiardja <reubendb at > utk.edu>wrote: > >> Hello, >> >> I would like to compare / benchmark several different PETSc installations >> on my >> system. The differences may be: different compilers or compiler flags, >> different >> libraries used by PETSc (e.g. BLAS&LAPACK, and any other packages used by >> PETSc that I included in the build like SuperLU_Dist, Metis/ParMetis, >> Hypre, >> etc). Some of these libraries were provided by different vendors and some >> we >> built ourselves and I'd like to compare the performances as used by PETSc. >> >> So I am wondering if there's any kind of general PETSc benchmarking code >> out >> there that can do this. The more thorough parts of PETSc that gets >> exersized >> by the benchmark code the better, but even if only parts are tested it'd >> be >> useful too. Also I'm more interested on the benchmark code that can run >> with >> larger number of processors (e.g. order of hundreds and thousands). I've >> tried >> Google search but that hasn't yielded much of anything. So any help in >> pointing me to the right direction is greatly appreciated. >> >> Thank you >> RDB >> >> >> >> >> >> >> >> >> >> > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20121109/aa8a347a/attachment.html>
