EXECUTION_COMMAND_LINE (mkdir) error

[Thor Tepper García]

Hi all,

I have recently been experiencing a weird error. It consists in the code

*occasionally* failing to create a new output subdirectory via the Fortran subroutine

EXECUTION_COMMAND_LINE. The relevant snippet is around lines 39-43 is in the old(er) version

of output_amr.f90.

This has been happening since I started running jobs on a new cluster, therefore I strongly

suspect that it is not a bug in the code but rather has to do with the something particular to

the cluster, e.g. the openmpi version (2.1) or something else.

As mentioned above, this only happens occasionally, i.e. the code runs fine, writing output

to disk, until it suddenly fails for no apparent reason. My jobs allocate enough memory,

and there is more than enough memory on disk for the output to be written. I know this

because restarting the job generally fixes the issue, albeit only temporarily, i.e. the job fails again

at a later stage.

Looking for some pointers in the relevant forums (or fora?) all I could gleaned is that

EXECUTION_COMMAND_LINE may fail to create a new directory if there is not enough

*virtual* memory available in the system; which really does not help much.

So my questions are:

1) Has anyone experienced this issue before?

2) Does anyone have an idea where to start looking for a possible solution or at least a workaround?

Needless to say that I have contacted the help team responsible for the cluster (still waiting for their reply).

Cheers,

Thor.

[Asger Gronnow]

Hi Thor,
I've experienced this problem before on multiple machines. My workaround is to set withoutmkdir to .true. inside amr_parameters.f90 such that Ramses doesn't try to create the directories. Then you have to create them manually in the output directory before running the simulation with a command such as: for i in $(seq -f "%05g" 1 100); do mkdir output_$i;done

Cheers,
Asger