[petsc-users] hypre support
Dario Isola
dario.isola at newmerical.com
Fri May 16 11:49:40 CDT 2014
Thanks a lot for your answers.
I ran it with
-ksp_type gmres -pc_type hypre -pc_hypre_type euclid
and it worked very well. Thanks.
I then tried to use boomeramg as a preconditioner coupled with
Richardson but I was not successful, it failed to solve the system and
returned nans.
-ksp_type richardson -pc_type hypre -pc_hypre_type boomeramg
-pc_hypre_boomeramg_relax_type_all SOR/Jacobi
-pc_hypre_boomeramg_print_debug -ksp_view -ksp_monitor_true_residual
and i got the following
===== Proc = 0 Level = 0 =====
Proc = 0 Coarsen 1st pass = 0.000000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 18308 nc_offd = 0
Proc = 0 iter 2 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 7 nc_offd = 0
===== Proc = 0 Level = 1 =====
Proc = 0 Coarsen 1st pass = 0.010000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 8725 nc_offd = 0
Proc = 0 iter 2 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 16 nc_offd = 0
===== Proc = 0 Level = 2 =====
Proc = 0 Coarsen 1st pass = 0.000000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 4721 nc_offd = 0
Proc = 0 iter 2 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 13 nc_offd = 0
Proc = 0 iter 3 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 4 nc_offd = 0
===== Proc = 0 Level = 3 =====
Proc = 0 Coarsen 1st pass = 0.000000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 2495 nc_offd = 0
Proc = 0 iter 2 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 22 nc_offd = 0
Proc = 0 iter 3 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 4 nc_offd = 0
===== Proc = 0 Level = 4 =====
Proc = 0 Coarsen 1st pass = 0.000000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 1337 nc_offd = 0
Proc = 0 iter 2 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 13 nc_offd = 0
Proc = 0 iter 3 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 2 nc_offd = 0
===== Proc = 0 Level = 5 =====
Proc = 0 Coarsen 1st pass = 0.000000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 695 nc_offd = 0
Proc = 0 iter 2 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 3 nc_offd = 0
===== Proc = 0 Level = 6 =====
Proc = 0 Coarsen 1st pass = 0.000000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 343 nc_offd = 0
Proc = 0 iter 2 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 21 nc_offd = 0
Proc = 0 iter 3 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 2 nc_offd = 0
===== Proc = 0 Level = 7 =====
Proc = 0 Coarsen 1st pass = 0.000000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 174 nc_offd = 0
Proc = 0 iter 2 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 15 nc_offd = 0
Proc = 0 iter 3 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 2 nc_offd = 0
===== Proc = 0 Level = 8 =====
Proc = 0 Coarsen 1st pass = 0.000000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 81 nc_offd = 0
Proc = 0 iter 2 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 13 nc_offd = 0
===== Proc = 0 Level = 9 =====
Proc = 0 Coarsen 1st pass = 0.000000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 37 nc_offd = 0
Proc = 0 iter 2 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 6 nc_offd = 0
===== Proc = 0 Level = 10 =====
Proc = 0 Coarsen 1st pass = 0.000000
Proc = 0 Coarsen 2nd pass = 0.000000
Proc = 0 Initialize CLJP phase = 0.000000
Proc = 0 iter 1 comm. and subgraph update = 0.000000
Proc = 0 CLJP phase = 0.000000 graph_size = 11 nc_offd = 0
0 KSP preconditioned resid norm 7.299769365830e+14 true resid
norm 8.197927963033e-03 ||r(i)||/||b|| 1.000000000000e+00
1 KSP preconditioned resid norm 2.319459389445e+28 true resid
norm 6.152576199945e+12 ||r(i)||/||b|| 7.505038136086e+14
KSP Object: 1 MPI processes
type: richardson
Richardson: damping factor=1
maximum iterations=90, initial guess is zero
tolerances: relative=0.1, absolute=1e-50, divergence=100000
left preconditioning
using PRECONDITIONED norm type for convergence test
PC Object: 1 MPI processes
type: hypre
HYPRE BoomerAMG preconditioning
HYPRE BoomerAMG: Cycle type V
HYPRE BoomerAMG: Maximum number of levels 25
HYPRE BoomerAMG: Maximum number of iterations PER hypre call 1
HYPRE BoomerAMG: Convergence tolerance PER hypre call 0
HYPRE BoomerAMG: Threshold for strong coupling 0.25
HYPRE BoomerAMG: Interpolation truncation factor 0
HYPRE BoomerAMG: Interpolation: max elements per row 0
HYPRE BoomerAMG: Number of levels of aggressive coarsening 0
HYPRE BoomerAMG: Number of paths for aggressive coarsening 1
HYPRE BoomerAMG: Maximum row sums 0.9
HYPRE BoomerAMG: Sweeps down 1
HYPRE BoomerAMG: Sweeps up 1
HYPRE BoomerAMG: Sweeps on coarse 1
HYPRE BoomerAMG: Relax down SOR/Jacobi
HYPRE BoomerAMG: Relax up SOR/Jacobi
HYPRE BoomerAMG: Relax on coarse Gaussian-elimination
HYPRE BoomerAMG: Relax weight (all) 1
HYPRE BoomerAMG: Outer relax weight (all) 1
HYPRE BoomerAMG: Using CF-relaxation
HYPRE BoomerAMG: Measure type local
HYPRE BoomerAMG: Coarsen type Falgout
HYPRE BoomerAMG: Interpolation type classical
linear system matrix = precond matrix:
Matrix Object: 1 MPI processes
type: seqbaij
rows=22905, cols=22905, bs=5
total: nonzeros=785525, allocated nonzeros=785525
total number of mallocs used during MatSetValues calls =0
block size is 5
Do you guys have any suggestion? Is it possible that I am haven't
initialized boomeramg properly? Or it is just my system equations that
can not be solved by AMG?
Sincerely,
Dario
On 05/16/2014 11:54 AM, Barry Smith wrote:
> On May 16, 2014, at 10:46 AM, Dario Isola <dario.isola at newmerical.com> wrote:
>
>> Dear all,
>>
>> I am investigating the use of hypre+petsc. I was able to successfully configure, install, compile petsc 3.3 with the external package for hypre.
>>
>> I tried to run it with the following options
>> -pc_type hypre -pc_type_hypre pilut -ksp_type richardson
>> and, although he did not complain, it does not solve the system either.
> Do you meaning it did not converge? At first always run with -ksp_view (or -snes_view if using snes or -ts_view if using ts) and -ksp_monitor_true_residual to see what is going on.
>
>> -pc_type_hypre pilut
> is wrong it is -pc_hypre_type pilut
>
> Note that pilut will generally not work with Richardson you need a “real” Krylov method like GMRES.
>
> Also the ilu type preconditioners don’t scale particularly well though occasionally they can be fine.
>
>> To what extent is hypre supported by petsc? More specifically, what kind of matrices?
> If it cannot handle the matrix type it would give an error message. Hypre uses a format like AIJ so you should use AIJ. Note that you can make the matrix type a runtime option so you don’t have to compile in that it is BAIJ.
>
>
>> I am using a baij matrix.
>>
>> Thanks in advance,
>> D
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20140516/1925e267/attachment.html>
More information about the petsc-users
mailing list