[Dune] Parallel grid partitioning

Nagaiah Chamakuri nagaiah.chamakuri at gmail.com
Sat Oct 15 13:14:39 CEST 2011


Dear All,

I have a question about load balance issue. I have a grid where we placed
many
elements at one region and the other part of the domain consists of uniform
grid
(adapted grid  using UGGRID in 3D and saved in gmsh format).  I am solving a
three coupled
reaction-diffusion system using pdelab subroutines.(using power grid
function space).

1) using UGGRID(tetrahedron elements) : poor load balance over increasing
the processors.
I can see that parallel efficiency only 30% on 8 procs comparing with
sequential run.
even worse on (20% efficiency) 32 procs. I understood that,  due to the poor
load balance of
the grid, the parallel efficiency goes down. Am I right here or made a
stupid mistake?
 Is there a way to play around to resolve the  load balance issue using
UGGRID.

I tried with ALUGRID as well(compiled using METIS), still I can see the
similar behavior
even though changing the parameters in alugrid.cfg file
(from 11( PartGraphKway) to 12(PartGraphRecursive)).
Any idea to improve the load balance issue?

2) second issue about comparison sequential run cpu times for the same
problem using UG and ALU: In this case using UGGRID is about twice faster
than using ALUGRID.
Even I can see a similar behavior for the nonoverlappingsinglephaseflow
problem in
dune-pdelab-howto/src/convection-diffusion.

I am quite surprised why ALUGRID is 50% slower than UG for pdelab sub
routines???
did I make a mistake during the compilation (I used -DNDEBUG flag and -O3
during the ALUGRID and DUNE compilation) or did you observe the similar
behavior?

Thank you in advance.

Best Regards
Nagaiah
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.dune-project.org/pipermail/dune/attachments/20111015/4ed54ffb/attachment.htm>


More information about the Dune mailing list