[petsc-dev] Seeming performance regression with GAMG

[Lawrence Mitchell]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 27/04/15 18:58, Tobin Isaac wrote:
> On Mon, Apr 27, 2015 at 04:06:30PM +0100, Lawrence Mitchell wrote:
>>> Dear all,
>>> 
>>> we recently noticed a slowdown when using GAMG that I'm trying
>>> to track down in a little more detail.  I'm solving an Hdiv-L2 
>>> "helmholtz" pressure correction using a schur complement.  I 
>>> precondition the schur complement with 'selfp', which morally
>>> looks like a normal Helmholtz operator (except in the DG
>>> space).  The domain is very anisotropic (a thin atmospheric
>>> shell), so getting round to trying Toby's column-based
>>> coarsening plugin is on the horizon but I haven't done it yet.
>>> 
>>> I don't have a good feel for exactly when things go worse, but
>>> here are two data points:
>>> 
>>> A recentish master (e4b003c), and master from 26th Feb
>>> (30ab49e4).  I notice in the former that MatPtAP takes
>>> significantly longer (full logs below), different coarsening
>>> maybe?  As a point of comparison, the PCSetup for Hypre takes
>>> ballpark half a second on the same operator.
>>> 
>>> I test with KSP ex6 (with a constant RHS):
>>> 
>>> Any ideas?
> While there may be other changes that have affected your
> performance, I see two things in your logs:
> 
> - The coarse matrix is much smaller (3 vs. 592).  The default
> coarse equations limit was recently changed from 800 to 50.  You
> can recover the old behavior with `-pc_gamg_coarse_eq_limit 800`. -
> GAMG now uses the square of the adjacency graph only on the finest 
> level.  This means that matrices on the coarser levels will be 
> larger and have more entries, which probably explains the extra
> PtAP time.  Maybe Mark can explain the decision to make this
> change.

Yes, thanks, this is the change.  If I switch back to using the square
of the adjacency graph on all levels then I recover the previous
behaviour.  Mark, any comments?

Cheers,

Lawrence

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1

iQEcBAEBAgAGBQJVP1ipAAoJECOc1kQ8PEYve1QH/3SChuuIZhDOoOOfxYnezCuC
beXRvqrjem6SWMBIft67kvmAr9htGsiGg/vE+NRP9EwNu20kD+fYocJ+mBBcxFit
2FdUperqA8F9gys2wJ+TijEQx58ieBNJJyPmQRa1DOmKHmD/WWcCIBV5q+yML9kt
B3z59ujUH1kJrmmpYmUk6aMUDAG0qU2KlVHX+p6zL1kV7ooGdYcKbiAbYdYrKO1l
6vNw65+51Oiyi2X/oAsnxeJeAvMN2ljaSYLZJ1RZgAndSLi6bcZPQZvtzwXfOLs9
C0qSPfXbd5O2DoenYnKzCS7t/kLnl0J3QimYex4tOSQxNQ7MtMG4CMNDmAIepUs=
=OY/s
-----END PGP SIGNATURE-----