> "Mark F. Adams" <[email protected]> writes: >> There are many parameters, including stupid things like solver >> parameters and stupider things like bugs, that can kill an iterative >> solver and we are trying to figure out what is wrong here. You should >> be running with about 1/10 or 1/100, or even 1/1000, the number of >> processors that you are using here. If you eventually want to run in >> this regime then that is fine but for debugging it helps to be in a >> more normal regime, because then you (or us) can start to see >> secondary things like super slow flop rates that can provide a hint of >> what is going wrong. It is also better to start in serial so there >> are less moving parts > > Pierre, can you scale the problem down to 10k or 100k degrees of freedom > and make sure that you can reproduce the qualitative convergence > behavior. > > Write the matrix to a file using > > -ksp_view_mat binary:matrix-100k-jolivet > > then put the file matrix-100k-jolivet and matrix-100k-jolivet.info > somewhere we can find it or anonymous ftp to ftp.mcs.anl.gov, cd > incoming, and upload the file. > > This will let us experiment quickly. >
I really don't want to bother you more than I already have, but since you asked for it, I uploaded a P_1 3D Poisson matrix on ANL's ftp. On my desktop, using 2 processes (~70k unkowns/proc), AMG is 8x slower (setup ~8 seconds) than BoomerAMG or GASM (setup ~1 second). Please feel free to drop the matter, I'll make some more runs at larger scale and just pick the best results. Thanks again for your help.
