Justin is right that parallelism will be of limited value for such small systems. This looks like a serial optimization job. Moreover, in this case, a better numerical method usually trumps any kind of machine optimization.
Matt On Fri, Oct 20, 2017 at 2:55 AM, Justin Chang <[email protected]> wrote: > 600 unknowns is way too small to parallelize. Need at least 10,000 > unknowns per MPI process: https://www.mcs.anl.gov/petsc/documentation/faq. > html#slowerparallel > > What problem are you solving? Sounds like you either compiled PETSc with > debugging mode on or you just have a really terrible solver. Show us the > output of -log_view. > > On Fri, Oct 20, 2017 at 12:47 AM Luca Verzeroli < > [email protected]> wrote: > >> Good morning, >> For my thesis I'm dealing with GALILEO, one of the clusters owned by >> Cineca. http://www.hpc.cineca.it/hardware/galileo >> The first question is: What is the best configuration to run petsc on >> this kind of cluster? My code is only a MPI program and I would like to >> know if it's better to use more nodes or more CPUs with mpirun. >> This question comes from the speed up of my code using that cluster. I >> have a small problem. The global matrices are 600x600. Are they too small >> to see a speed up with more mpiprocess? I notice that a single core >> simulation and a multi cores one take a similar time (multi core a second >> more). The real problem comes when I have to run multiple simulation of the >> same code changing some parameters. So I would like to speed up the single >> simulation. >> Any advices? >> >> >> Luca Verzeroli >> > -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener https://www.cse.buffalo.edu/~knepley/ <http://www.caam.rice.edu/~mk51/>
