Ticket 6395

One of our system admins is experimenting with setting up a BeeOND file system (https://www.beegfs.io/wiki/BeeOND) in the job prolog (and tearing it down in the epilog). The trouble that he's running into is that only one node in the allocation sets up the metadata server for the BeeOND file system. The prolog script takes longer to run on that node, and so tasks can get launched on other nodes before the file system is fully set up and available. I was looking for a Slurm configuration option that would prevent any tasks from launching until the prolog has completed on all of the nodes in the allocation. 'PrologFlags=Alloc' appears to do what I want for jobs launched with sbatch or salloc, but if I launch directly with srun, I still see asynchronous launch behavior.

Here's a simple reproducer:

I have a prolog script that sleeps for 60s on the first node in an allocation, and 30s on all other nodes:

[day36@opal186:prolog_test]$ cat waiter.sh
#!/bin/sh

thishost=`/bin/hostname -s`
firsthost=`/bin/scontrol show hostnames ${SLURM_NODELIST} | head -1`

if [[ ${SLURMD_NODENAME} = ${firsthost} ]]; then
	echo "first!" > /g/g0/day36/waitout.${thishost}
	sleep 60
else
	sleep 30
fi

exit 0
[day36@opal186:prolog_test]$

I have a script that prints the host and date:

[day36@opal186:prolog_test]$ cat checker.sh
#!/bin/sh

hostname
date
[day36@opal186:prolog_test]$

and I have a batch script that runs that checker script on two nodes:

[day36@opal186:prolog_test]$ cat check_stuff.sbatch 
#!/bin/sh

#SBATCH -N 2
#SBATCH --reservation=test

srun -N 2 --ntasks-per-node=1 checker.sh
[day36@opal186:prolog_test]$

If I have the waiter.sh prolog script in place and no PrologFlags, one checker.sh task runs 30s before the other one:

[day36@opal186:prolog_test]$ date
Wed Jan 23 09:40:07 PST 2019
[day36@opal186:prolog_test]$ srun -N2 -n2 --reservation=test checker.sh 
opal109
Wed Jan 23 09:40:40 PST 2019
opal108
Wed Jan 23 09:41:08 PST 2019
[day36@opal186:prolog_test]$

If I have the waiter.sh prolog script in place, PrologFlags=Alloc, and run in an sbatch script, the tasks run at the same time:

[day36@opal186:prolog_test]$ sbatch check_stuff.sbatch 
Submitted batch job 46764
[day36@opal186:prolog_test]$ date
Wed Jan 23 10:08:32 PST 2019
…
 [day36@opal186:prolog_test]$ cat slurm-46764.out 
opal108
Wed Jan 23 10:09:32 PST 2019
opal109
Wed Jan 23 10:09:32 PST 2019
[day36@opal186:prolog_test]$

But, if I have the waiter.sh prolog script in place, PrologFlags=Alloc, and run directly with srun, one task still runs 30s before the other:

[day36@opal186:prolog_test]$ date
Wed Jan 23 09:43:43 PST 2019
[day36@opal186:prolog_test]$ srun -N2 -n2 --reservation=test checker.sh 
opal109
Wed Jan 23 09:44:17 PST 2019
opal108
Wed Jan 23 09:44:44 PST 2019

Is there a way to get task launch to wait for all prologs to complete, or do we have to write that synchronization into the prolog scripts themselves?