On Thu, Mar 30, 2017 at 3:04 PM, Justin Chang <[email protected]> wrote:
> Yeah based on my experiments it seems setting pc_mg_levels to $DAREFINE + > 1 has decent performance. > > 1) is there ever a case where you'd want $MGLEVELS <= $DAREFINE? In some > of the PETSc tutorial slides (e.g., http://www.mcs.anl.gov/ > petsc/documentation/tutorials/TutorialCEMRACS2016.pdf on slide 203/227) > they say to use $MGLEVELS = 4 and $DAREFINE = 5, but when I ran this, it > was almost twice as slow as if $MGLEVELS >= $DAREFINE > Depending on how big the initial grid is, you may want this. There is a balance between coarse grid and fine grid work. > 2) So I understand that one cannot refine further than one grid point in > each direction, but is there any benefit to having $MGLEVELS > $DAREFINE by > a lot? > Again, it depends on the size of the initial grid. On really large problems, you want to use GAMG as the coarse solver, which will move the problem onto a smaller number of nodes so that you can coarsen further. Matt > Thanks, > Justin > > On Thu, Mar 30, 2017 at 2:35 PM, Barry Smith <[email protected]> wrote: > >> >> -da_refine $DAREFINE determines how large the final problem will be. >> >> By default if you don't supply pc_mg_levels then it uses $DAREFINE + 1 >> as the number of levels of MG to use; for example -da_refine 1 would result >> in 2 levels of multigrid. >> >> >> > On Mar 30, 2017, at 2:17 PM, Justin Chang <[email protected]> wrote: >> > >> > Hi all, >> > >> > Just a general conceptual question: say I am tinkering around with SNES >> ex48.c and am running the program with these options: >> > >> > mpirun -n $NPROCS -pc_type mg -M $XSEED -N $YSEED -P $ZSEED >> -thi_mat_type baij -da_refine $DAREFINE -pc_mg_levels $MGLEVELS >> > >> > I am not too familiar with mg, but it seems to me there is a very >> strong correlation between $MGLEVELS and $DAREFINE as well as perhaps even >> the initial coarse grid size (provided by $X/YZSEED). >> > >> > Is there a rule of thumb on how these parameters should be? I am >> guessing it probably is also hardware/architectural dependent? >> > >> > Thanks, >> > Justin >> >> > -- 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
