[Dune] Parallel CG Solver

[Arne Rekdal]

Hello again!

Thank you for all your help. I have now got a program that works for  
two processors. When I try to use more than two processors the program  
crashes.

The crucial function call is
  u.fillIndexInfoFromGrid(indexinfo)

where u is a LeafP1Function. I have tried OpenMPI and MPICH libraries,  
with the same result. Output from the run

tr-lcdfe:~/dune:$ mpirun -np 4 dune-arne/dune_parallel_matrix 100
Program is using 4 processes.
Rank: 0 Subdomain contains 2704 vertices.
Rank: 2 Subdomain contains 2704 vertices.
Rank: 1 Subdomain contains 2704 vertices.
Rank: 3 Subdomain contains 2704 vertices.
0 - MPI_TEST : Message truncated
[0]  Aborting program !
[0] Aborting program!
p2_4253:  p4_error: : 14
rm_l_2_4255: (0.199219) net_send: could not write to fd=5, errno = 32
p0_4166:  p4_error: : 14
1 - MPI_TEST : Message truncated
[1]  Aborting program !
[1] Aborting program!

Is there anyone that could test if this is an issue with my setup or  
an error in my code?

I am also not satisfied with the scaling of runtime vs. problem size  
on two processors. If I halve my grid spacing then my problem size  
increases with a factor of four (in 2D). I went from 13 sec to 320 sec  
spent in CGSolver by going from 100 to 200 nodes in each direction  
(resp. 6 iterations and 8 iterations used).

Thank you!


Best regards
Arne Rekdal

-------------- next part --------------
A non-text attachment was scrubbed...
Name: dune_parallel_matrix.cc
Type: text/x-c++
Size: 10840 bytes
Desc: not available
URL: <https://lists.dune-project.org/pipermail/dune/attachments/20081204/c5f74841/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: laplace_arne.hh
Type: text/x-c++hdr
Size: 7590 bytes
Desc: not available
URL: <https://lists.dune-project.org/pipermail/dune/attachments/20081204/c5f74841/attachment.hh>