[petsc-users] Multigrid iterations for different versions of Petsc

[Barry Smith]

  First make sure both runs think they are doing the exact same solver options so run both with -ksp_view and make sure the output is the same.

  I note that the 3.5 run has

KSPSolve            2200 1.0 8.3100e+00 1.0 7.28e+08 1.0 1.4e+05 2.5e+03 5.9e+03 55 67 45 44 79  83 83 47 80100   700
PCApply             5764 1.0 9.4324e-01 1.1 3.12e+08 1.0 2.5e+04 6.4e+01 0.0e+00  6 29  8  0  0   9 36  8  0  0  2646

while the 3.4 has

KSPSolve            4125 1.0 7.7953e+00 1.0 1.39e+09 1.0 2.6e+05 2.5e+03 1.1e+04 60 65 45 53 82  82 83 47 80100  1424
PCApply            10769 1.0 4.4700e+00 1.0 5.80e+08 1.0 4.6e+04 6.4e+01 0.0e+00 34 27  8  0  0  47 35  8  0  0  1038

why twice as many KSPSolves? Are you calling the KSPSolve more ?

  Barry

On Oct 6, 2014, at 7:23 AM, Filippo Leonardi <filippo.leonardi at sam.math.ethz.ch> wrote:

> Hi,
> 
> Again me with multigrid. Sorry for the annoyance, but I am a bit worried about 
> my multigrid solver: I still can't figure out where (if something) my code is 
> doing wrong.
> 
> I post the details of a run with my local machine and Petsc 3.5 compared with 
> the runs done with Petsc 3.4.3 on a cluster (the most recent version 
> available).
> 
> The problem is that the run in the cluster (3.4.3) does twice many iteration 
> per krylow solver (as the local 3.5) and goes even worse when increasing the 
> mesh. Also, while on my local machine I get 7 iterations (which is quite 
> nice), on the cluster I get 14 iteration, which is not what I expect from 
> multigrid.
> 
> Are there been any changes on the MG solver that can influence the behaviour of 
> the solver?
> 
> http://www.mcs.anl.gov/petsc/documentation/changes/35.html
> seems not to include anything relevant.
> 
> Everything is the same: outer iteration with gmres, lu as coarse solver etc 
> (all default), I just adapted the APIs. Thus I'd expect more or less the same 
> result.
> 
> Best,
> Filippo
> 
> I attach the profiles.<3-4.txt><3-5.txt><ETHZ.vcf>