chain stop when MPI communicating about cosmomc HOT 10 CLOSED

It's odd if it only happens after a long time. Which compiler are you using? If ifort 2017.0.4, then get it updated - that version has bugs.

from cosmomc.

Thank you @cmbant
I am using gcc-6.4.1
Is there any way to modify some functions in the code to fit the gcc? Our cluster is IBM and will not use ifort...
Thanks

from cosmomc.

It should work find in gcc, though it's possible a compiler bug has been introduced.

from cosmomc.

Please try ad10c52 (though that fix was specific to resuming from checkpoint files, it sounds like the issue).

from cosmomc.

Hi! @cmbant
Just try the new sources over a IBM cluster, the same errors are still happened.

from cosmomc.

Make sure you have deleted all old chain .chk files, they may be corrupted.

from cosmomc.

@cmbant I just tried, but the error happen as before. Sorry to disturb you again.

from cosmomc.

Do you chains have more than 500000 samples when this happens, and if so does it happen after thin() is called (in ObjectLists.f90)?

from cosmomc.

@cmbant Sorry to disturb you again.
The chain steps = 59101 for example, and I have tested over the git cloned version just now, by testing the generic_samplers. The *.chk files are generated when I pressed Ctrl-C but restart still failed and showing
1 Reading checkpoint from chains/test2_1.chk
TObjectList: Unknown type to save

Looks different than before but still don not work.
Thank you

from cosmomc.

This seems to be a version-specific gfortran compiler bug see
https://cosmocoffee.info/viewtopic.php?f=11&t=2827

from cosmomc.