Change default /tmp/coi_procs location? (COI_PROCS_DIR)

[Chris_Samuel]

Hi there,

Our Intel Xeon Phi cards are in diskless compute nodes in our IBM iDataplex cluster and we have an issue where /tmp gets filled up due to files being left behind in the /tmp/coi_procs directory.   Ideally we would like that location to be determined by $TMPDIR as then they will end up on our high performance scratch GPFS filesystem with 254TB free, rather than the 1GB of RAM we allocate to /tmp on the nodes. :-)

Is that possible please?

thanks,
Chris

[Sumedh_N_Intel]

Just to be sure I am understanding everything correctly, are you refering to /tmp/coi_procs on the host or on the coprocessor? Also, based on my understanding /tmp/coi_procs files are deleted as soon your application exits (i.e. when there are no active coi processes). I am not sure, what is causing the files to be left behind in /tmp/coi_procs.

Let me take a closer look and I will get back to you.  

[Chris_Samuel]

Hi Sumedh, This is /tmp/coi_procs on the host, it does seem odd that they are not getting tidied up but that does appear to be the case. Thanks! Chris

[Sumedh_N_Intel]

Hi Chris, 

These files are normally left behind when the host program crashes. The files are for the profiler and the debugger. If no COI processes are running, the user is free to delete these files. In the Intel MPSS 3.1 release, the default file generation is off and hence the user won't have to worry about this. 

I hope this helps. 

[Chris_Samuel]

Hi Sumedh,

Thanks for that, sounds great.  I suspect in this case the simulations hit their walltime rather than crashing and so were terminated by SLURM and didn't get a chance to clean up after themselves.

All the best,
Chris