[slurm-users] Problem launching interactive jobs using srun

[Pickering, Roger (NIH/NIAAA) [E]]

I'm confused.  Why would you want to run an interactive program using srun? 

Roger

-----Original Message-----
From: Andy Georges [mailto:Andy.Georges at UGent.be] 
Sent: Friday, March 09, 2018 12:20 PM
To: slurm-users at lists.schedmd.com
Subject: [slurm-users] Problem launching interactive jobs using srun

Hi,


I am trying to get interactive jobs to work from the machine we use as a login node, i.e., where the users of the cluster log into and from where they typically submit jobs.


I submit the job as follows:

vsc40075 at test2802 (banette) ~> /bin/salloc -N1 -n1 /bin/srun bash -i
salloc: Granted job allocation 41
salloc: Waiting for resource configuration
salloc: Nodes node2801 are ready for job

…
hangs


On node2801, the slurmd log has the following information:


[2018-03-09T18:16:08.820] _run_prolog: run job script took usec=10379 [2018-03-09T18:16:08.820] _run_prolog: prolog with lock for job 41 ran for 0 seconds [2018-03-09T18:16:08.829] [41.extern] task/cgroup: /slurm/uid_2540075/job_41: alloc=800MB mem.limit=800MB memsw.limit=880MB [2018-03-09T18:16:08.830] [41.extern] task/cgroup: /slurm/uid_2540075/job_41/step_extern: alloc=800MB mem.limit=800MB memsw.limit=880MB [2018-03-09T18:16:11.824] launch task 41.0 request from 2540075.2540075 at 10.141.21.202 (port 61928) [2018-03-09T18:16:11.824] lllp_distribution jobid [41] implicit auto binding: cores,one_thread, dist 1 [2018-03-09T18:16:11.824] _task_layout_lllp_cyclic [2018-03-09T18:16:11.824] _lllp_generate_cpu_bind jobid [41]: mask_cpu,one_thread, 0x1 [2018-03-09T18:16:11.834] [41.0] task/cgroup: /slurm/uid_2540075/job_41: alloc=800MB mem.limit=800MB memsw.limit=880MB [2018-03-09T18:16:11.834] [41.0] task/cgroup: /slurm/uid_2540075/job_41/step_0: alloc=800MB mem.limit=800MB memsw.limit=880MB [2018-03-09T18:16:11.836] [41.0] error: connect io: Connection refused [2018-03-09T18:16:11.836] [41.0] error: IO setup failed: Connection refused [2018-03-09T18:16:11.905] [41.0] _oom_event_monitor: oom-kill event count: 1 [2018-03-09T18:16:11.905] [41.0] error: job_manager exiting abnormally, rc = 4021 [2018-03-09T18:16:11.905] [41.0] error: _send_launch_resp: Failed to send RESPONSE_LAUNCH_TASKS: Connection refused [2018-03-09T18:16:11.907] [41.0] done with job


We are running slurm 17.11.4.


When I change to the same user on both the master node (running slurmctld) and worker nodes (running slurmd), things work just fine. I would assume I need not run slurmd on the login node for this to work?


Any pointers are appreciated,
— Andy