Segmentation fault with WET_DRY

[nganju]

Hello-
I have two computers: a dual-processor Linux with Intel 9, and a dual-core windowsXP/Cygwin with Intel 8.

The configuration that I have been running on the Linux machine for some time has been with WET_DRY off. When I turn WET_DRY on, I get a segmentation fault after the first time step (stack size is "unlimited").

Code: Select all

STEP    time[DAYS]  KINETIC_ENRG    POTEN_ENRG    TOTAL_ENRG  NET_VOLUME  trd
      0   274.729200  0.000000E+00  3.368108E+01  3.368108E+01  2.407436E+09  0
      DEF_HIS   - creating history file: ./ROMS/output/suisun_his_0001.nc
      WRT_HIS   - wrote history  fields (Index=1,1) into time record = 0000001
forrtl: severe (174): SIGSEGV, segmentation fault occurred
Image              PC        Routine            Line        Source
oceanO             08080CD0  Unknown               Unknown  Unknown

Stack trace terminated abnormally.

If I compile for a serial run (OpenMP off), the WET_DRY model runs with no error.

When I run the same model on the Cygwin machine, I have no problems, whether I compile with OpenMP on or off.

I read over previous posts on segmentation faults, this seems like the only one that has happened post-compilation, during the run (as far as I could tell). I have fooled with the partitioning and compilation flags to no avail, I wonder if this is an Intel 9 vs. 8 difference? Any ideas are appreciated...

[cvl]

My experience with several post-compile segmentation faults (at approximately the same point you experienced) was with Fedora 6 and Intel 9 on a 64 bit machine.

Said problems have abated with MPI since I've gone through one Intel upgrade and several kernel upgrades, but I don't believe I've gotten a good run out of OpenMP in a while. You might try boosting the tiling to very high numbers and seeing what that does?

The stack-size issue may be relevant, but the solution is apparently more complex than callling ulimit and futzing about with /etc/*rc files. Since I'm not in a hurry, I'm sitting on my hands until v3.0 arrives.

[schen]

Hi nganju,
I encountered a similar post-compilation segmentation fault using OpenMp. Have you resolved that problem yet? If so, do you mind sharing the trick? Thanks

Shih-Nan

[nganju]

Hello Shih-Nan-
I haven't figured it out yet, and I am first trying to get a complete run with sed_morph before i try switching wet_dry on again. When I do get to that point, I will probably just run it in serial, and do two runs at once on my machine...

[m.hadfield]

I've said it before and I'll say it again: one of the things you should do is to build the model with USE_DEBUG on and run the executable. If the segmentation fault occurs because of an array bounds overflow, this will catch it.