Everyone, I have added into branch master master commit 5272c31 the options
-pc_hypre_boomeramg_nodal_coarsen <n> - where n is from 1 to 6 (see HYPRE_BOOMERAMGSetNodal()) -pc_hypre_boomeramg_vec_interp_variant <v> where v is from 1 to 4 (see HYPRE_BoomerAMGSetInterpVecVariant()) to allow taking advantage of MatSetNearNullSpace() for hypre BoomerAMG You must pass both of these values. Try all combinations they produce very different performance (using 0 means ignore the near null space) Since hypre BoomerAMG is very good regardless of this additional information it is generally unlikely you will see much improvement. In particular if your null space is just the constants I would not expect improvement. But give it a try. Barry > On Jul 24, 2015, at 10:25 AM, Young, Matthew, Adam <[email protected]> wrote: > > Thanks for the feedback, everyone. My null space is simply the constant > vectors and I am setting MatSetNullSpace based on code I had been testing > with other PC/KSP options, which grew out of examples. GMRES/BoomerAMG seems > to be the best fit for my problem but I stumbled upon that via > experimentation (Hurray for KSP/PCSetFromOptions!), rather than via a more > informed analysis. I am still in the process of learning about both the GMRES > algorithm and multigrid in general, so I asked this question in case it > turned out that I could improve my working code. > > I'd be interested in that interface to HYPRE_BoomerAMGSetInterpVectors, if it > makes it into PETSc. > > --Matt > > -------------------------------------------------------------- > Matthew Young > Graduate Student > Boston University Dept. of Astronomy > -------------------------------------------------------------- > > > ________________________________________ > From: [email protected] [[email protected]] on > behalf of Barry Smith [[email protected]] > Sent: Thursday, July 23, 2015 6:13 AM > To: Lawrence Mitchell > Cc: [email protected] > Subject: Re: [petsc-users] Null space/Near null space > >> On Jul 23, 2015, at 3:42 AM, Lawrence Mitchell >> <[email protected]> wrote: >> >> -----BEGIN PGP SIGNED MESSAGE----- >> Hash: SHA1 >> >> On 22/07/15 23:57, Barry Smith wrote: >>> >>> If your matrix has a null space you should always be setting >>> MatSetNullSpace. >>> >>> If you know the near null space then you should also always set it >>> (it cannot do any harm) because some preconditioners can take >>> advantage of it. >>> >>> I do not think PETSc currently has code that transfers near null >>> space information to hypre; I do not know if hypre has any >>> algorithms that can take advantage of it. Only some AMG multigrid >>> methods can utilize near null space information. >> >> FWIW, the HYPRE manual suggests yes: >> >> If the block size is set (PETSc already does this) and nodal systems >> coarsening is requested (PETSc has an option for this) then one can >> provide the near null modes with HYPRE_BoomerAMGSetInterpVectors, >> passing an array of Vecs (PETSc does not do this last step). > > Thanks. So someone should make this small code addition plus a test case > that utilizes it to demonstrate that it functions and wrap that up in a pull > request. > > Barry > >> >> Lawrence >> -----BEGIN PGP SIGNATURE----- >> Version: GnuPG v1 >> >> iQEcBAEBAgAGBQJVsKjjAAoJECOc1kQ8PEYvWhQIALEe6KAnENO8yveClQQAUImw >> xMkHXkuHiYtnbPZhQH8YeuVEx71pkmimXc0wRk532ee0FjfQZypuKybKrpFtVSvT >> HsvpFJwE0n3HxwaVdIYYtG8c0c/Lj3326u1oFxCN5trbcUAY5VUSb41fYPGrLc6q >> UQmWwMYgqPBdsGOykdOkVGGvJFbcc2CdR99Of3sEu+C0ub53/8T2Cav4yxAG32rn >> +7lxX+a74nN+M6lRCGI8D9HiKLwKtxUKhO8W3+QeTA+y0Hf4ytZ/4AkiPScU7j0n >> KwQ5vBdIt9hK8Qc8RU28gAn7uS0Ih4CgsrTtKP4Fdd5ksMEWTRZBQAimd2Yd5Cw= >> =bHru >> -----END PGP SIGNATURE----- >
