I have found and ran example 10: http://slurm.schedmd.com/cpu_management.html#Example10. I tried running the example on two partitions, one partition with shared=no and one with shared=exclusive. Both partitions behave in the same way, i.e. all six tasks are run one a single CPU. I want the six tasks to run on different CPUs. The strange thing here is that according to all manuals I can find on SLURM, what I want should be the default setting. Hopefully anyone has a good idea what might be going on here. I have been struggling with this for days.
If it matters, all compute nodes run Ubuntu 14.04.2 LTS. 2015-07-16 17:36 GMT+02:00 Jeroen Meijer <[email protected]>: > Unlimited, but I don't see how this is relevant. I think I have found a > shorter example. The command: "srun --partition=m610 -N9 -n72 > --ntasks-per-node=8 --ntasks-per-socket=8 --ntasks-per-core=1 > --cpus-per-task=1 --exclusive run-task.sh" runs 8 tasks on 9 nodes, but all > tasks are executed on a single core. Any idea why? > > 2015-07-16 15:24 GMT+02:00 Morris Jette <[email protected]>: > >> How much memory is each job allocated? >> >> On July 16, 2015 5:43:33 AM PDT, Jeroen Meijer <[email protected]> >> wrote: >>> >>> We are running a slurm controller (2.6.5) with built-in scheduler. No >>> matter which options I give to sbatch and srun I can only manage to run >>> multiple tasks on a single core. >>> >>> I have thousands of independent tasks I want to run. I should be able to >>> run them individually on a single core, right? I don't care about memory >>> bandwidth. I do care about using a dedicated core for each task. >>> >>> All the compute nodes have 8 cores and I want to run 8 tasks on a >>> dedicated core. So task 1 should run on core 1 and ... and task 8 should >>> run on core 8. What happens is that all 8 tasks are run on core 1. I do not >>> want this. I did also experiment with --exclusive and --shared. The used >>> partition is set in exclusive mode. >>> >>> Here is an example batch script I use: >>> #!/bin/bash >>> #SBATCH --partition=m610 -N9 >>> --output=~/experiments/scripts/slurm-out.log --open-mode=append >>> --cpus-per-task=1 --ntasks-per-core=1 --ntasks-per-node=8 >>> #steps 1 - 500 >>> srun -n1 -N1 --exclusive --time=35 >>> ~/experiments/scripts/steps/step_718f5c96-18da-421d-840a-ee94d4ddee18.sh & >>> ... thousands more similar tasks ... >>> >>> The full list of scheduling options is: >>> # SCHEDULING >>> #DefMemPerCPU=0 >>> FastSchedule=1 >>> #MaxMemPerCPU=0 >>> #SchedulerRootFilter=1 >>> #SchedulerTimeSlice=30 >>> SchedulerType=sched/builtin >>> SchedulerPort=7321 >>> SelectType=select/cons_res >>> SelectTypeParameters=CR_Core_Memory >>> SchedulerParameters=defer >>> >>> Any ideas what I am doing wrong? >>> >> >> -- >> Sent from my Android device with K-9 Mail. Please excuse my brevity. >> > >
