Hi Etienne,
for me, it is very difficult to make sense of speedup times on today's
multi-core CPUs. In addition, for the AMG solver, I am not so surprised,
because there is no parallel overhead at all for one core. The speedup
from two to four cores seems very reasonable. In order to go into more
detail, you should do the following:
- Set the verbosity level of the linear solver to 2 by, e.g., providing
[LinearSolver]
Verbosity = 2
in your .input file. Send the complete output of the runs to this list.
- Report/compare the times that are needed by the linear solvers only,
as given in the lines "Assemble/solve/update time:" of the output.
Comparison is only fair for the same number of Newton steps there.
- Do you use SuperLU for the coarse solves? If not, you might want to
apply istl-2.3.0.patch in the patches/ directory to dune-istl in order
to reduce the tolerance of the iterative coarse grid solver.
I would recommend the AMG for all parallel runs. As you experienced, the
other Dumux parallel solvers get into trouble already for very moderate
process numbers. Since we are not able to maintain and definitely not
able to improve these solvers, we most likely will get rid of them in
the not-to-distant future.
Kind regards
Bernd
On 10/21/2014 03:07 PM, Etienne Ahusborde wrote:
Dear DuMux developers,
With my colleagues we are working on two-phase multicomponent flow
with reactive transport and application to CO2 storage.
The problem is decoupled into two subproblems. The first one is
devoted to a two-phase two-component flow (H2O-CO2) while the second
one is devoted to a one-phase multicomponent reactive transport
problem for the other components.
We have successfully performed sequential simulations. No we would
like to perform parallel computations but we have some difficulties to
chose . In a previous mail, Bernd told to a DUMUX user that for your
parallel computations, you used AMG solver. Consequently we performed
some tests to compare the performances of the AMG solver in comparison
with the solver by default and we found a strange behaviour.
To explain our problem, we have installed the last version 2.6 of
DUMUX. We use the CO2 model on a 3D mesh (80000 elements). The problem
deals with injection of CO2. We use ALUGRID for the mesh. We want to
do only one iteration (DtInitial = 86400, TEnd = 86400) and compare
the time of simulation on 1, 2, 4, 8,... processors. The times of
simulation are displayed below:
#procs bicg amg
1 212.089 315.145
2 118.446 401.046
4 292.873 244.24
6 not done Stopped coarsening because of rate breakdown
For the bicg, with 4 processors, Newton solver does not converge with
Dt=86400 and the simulation needs 4 iterations to reach Tend so the
computation is slower than for one processor.
For the amg solver, the behaviour is even more strange since the
computations is faster with one processor than with two processors.
With 6 processors, the computation stopped but I think that it is due
to the small size of our mesh.
We expected that the amg solver will be faster that the bicg but it is
not the case.
Have you already found the same behaviour with your simulation? Is it
normal in your opinion? Which solver do you advise us?
We have also seen in the mail announcing the DUMUX 2.6 release that a
new direct solver UMFPACK is available. We would like to know how it
is possible to test it.
Thanks in advance.
Best regards
Etienne
_______________________________________________
Dumux mailing list
Dumux@listserv.uni-stuttgart.de
https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
--
_______________________________________________________________
Bernd Flemisch phone: +49 711 685 69162
IWS, Universität Stuttgart fax: +49 711 685 60430
Pfaffenwaldring 61 email: be...@iws.uni-stuttgart.de
D-70569 Stuttgart url: www.hydrosys.uni-stuttgart.de
_______________________________________________________________
_______________________________________________
Dumux mailing list
Dumux@listserv.uni-stuttgart.de
https://listserv.uni-stuttgart.de/mailman/listinfo/dumux