Attached are some results from scaling runs on Vulcan that I posted on another thread. The results are before and after fixing a problem with VecNorm.
And the second set of plots shows a comparison with the pure MPI case and MPI + pthreads. In both cases, the runs used two threads or ranks per core. Dave ________________________________________ From: Jed Brown [[email protected]] Sent: Monday, November 11, 2013 12:09 PM To: Mark Adams Cc: Nystrom, William D; For users of the development version of PETSc Subject: Re: [petsc-dev] Mira Mark Adams <[email protected]> writes: > I'm not sure if they really do have threads but they just wanted to see if > PETSc could use threads. It looks like only 40% of the run time is in > PETSc so we are shelving this but I wanted to try threads anyway. Okay, some programming is needed to make it function correctly on BG/Q. If you just want to get a rough sense, you could try the OpenMP branch >From the folks at Imperial.
<<attachment: KSPSolve_Time_vs_Node_Count_CPU_Pthread_6400_vulcan.png>>
<<attachment: KSPSolve_Time_vs_Node_Count_CPU_Pthread_MPI_6400_vulcan.png>>
