On Fri, Sep 12, 2014 at 11:11 AM, Jed Brown <[email protected]> wrote: > Rodrigo Broggi <[email protected]> writes: > >> Hey, I've copied the files from the site and normally run them ( without >> mpi) these are the results: > > You have to compare to parallel execution, either with threads or MPI > processes. The version in the PETSc source does that automatically. > What you'll (probably) see is that adding more threads/processes does > not improve the bandwidth.
If you manage to run with MPI up to 4 procs, here are the raw numbers from my workstation for reference. Number of MPI processes 1 Function Rate (MB/s) Copy: 7545.4086 Scale: 7761.3906 Add: 10409.8548 Triad: 10088.5244 Number of MPI processes 2 Function Rate (MB/s) Copy: 14150.5248 Scale: 14618.2789 Add: 19814.6343 Triad: 19591.9426 Number of MPI processes 3 Function Rate (MB/s) Copy: 21084.6303 Scale: 21615.8684 Add: 27826.3326 Triad: 27994.4276 Number of MPI processes 4 Function Rate (MB/s) Copy: 26555.4193 Scale: 26461.5099 Add: 31634.6075 Triad: 32217.4097 -- John ------------------------------------------------------------------------------ Want excitement? Manually upgrade your production database. When you want reliability, choose Perforce Perforce version control. Predictably reliable. http://pubads.g.doubleclick.net/gampad/clk?id=157508191&iu=/4140/ostg.clktrk _______________________________________________ Libmesh-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/libmesh-users
