Intro to SLURM slides

The Computer Science Department uses a “job scheduler” called SLURM. The purpose of a job scheduler is to allocate computational resources (servers) to users who submit “jobs” to a queue. The job scheduler looks at the requirements stated in the job's script and allocates to the job a server (or servers) which matches the requirements specified in the job script. For example, if the job script specifies that a job needs 192GB of memory, the job scheduler will find a server with at least that much memory free.

The job scheduler supports a direct login option (see below) that allows direct interactive logins to servers controlled by the scheduler, without the need for a job script.

• As of 10-Jan-2022, the SLURM job scheduler was updated to the latest revision which enforces memory limits. As a result, jobs that exceed their requested memory size will be terminated by the scheduler.

• As of 02-Apr-2022, the time limit enforcement policy within SLURM has changed. All jobs submitted with time limits are extended 60 minutes past the user-submitted time limit. E.g. If the user submits a job with a time limit of 10 minutes using parameter “-t 10”, SLURM will kill the job after 70 minutes.

• As of 23-Aug-2023, jobs on certain nodes are now preemptable. When using the sinfo or scontrol show partition command, any nodes within a partition named 'hipri' has preemption enabled. If you submit a job to one of these nodes, it may be suspended if the owner or research group submits a job to it as well.

• As of 26-Sep-2023, a default QoS was configured on all SLURM partitions in the CS cluster, limiting the number of concurrent jobs per user to 64.
  • Reservations may circumvent this limit by using the QoS csresnolim in srun commands or sbatch scripts using the parameter -q csresnolim or –qos=csresnolim see SLURM Reservations for more details

Slurm Commands Cheat Sheet

The SLURM commands below are ONLY available on the portal cluster of servers. They are not installed on the gpusrv* or the SLURM controlled nodes themselves.

To view information about compute nodes in the SLURM system, use the command sinfo.

[abc1de@portal04 ~]$ sinfo
PARTITION AVAIL  TIMELIMIT  NODES  STATE NODELIST
main*        up   infinite      3  drain falcon[3-5]
main*        up   infinite     27   idle cortado[01-10],falcon[1-2,6-10],lynx[08-12],slurm[1-5]
gpu          up   infinite      4    mix ai[02-03,05],lynx07
gpu          up   infinite      1  alloc ai04
gpu          up   infinite     12   idle ai[01,06],lynx[01-06],ristretto[01-04]

With sinfo we can see a listing of the job queues or “partitions” and a list of nodes associated with these partitions. A partition is a grouping of nodes, for example our main partition is a group of all general purpose nodes, and the gpu partition is a group of nodes that each contain GPUs. Sometimes hosts can be listed in two or more partitions.

To view jobs running on the queue, we can use the command squeue. Say we have submitted one job to the main partition, running squeue will look like this:

abc1de@portal01 ~ $ squeue
             JOBID PARTITION     NAME     USER    ST     TIME  NODES NODELIST(REASON)
            467039      main    my_job    abc1de  R      0:06      1 artemis1

and now that a node has been allocated, that node artemis1 will show as alloc in sinfo

abc1de@portal01 ~ $ sinfo
PARTITION     AVAIL  TIMELIMIT  NODES  STATE NODELIST
main*            up   infinite     37   idle hermes[1-4],artemis[2-7],slurm[1-5],nibbler[1-4],trillian[1-3],granger[1-6],granger[7-8],ai0[1-6]
main*            up   infinite      1  alloc artemis1
qdata            up   infinite      8   idle qdata[1-8]
qdata-preempt    up   infinite      8   idle qdata[1-8]
falcon           up   infinite     10   idle falcon[1-10]
intel            up   infinite     24   idle artemis7,slurm[1-5],granger[1-6],granger[7-8],nibbler[1-4],ai0[1-6]
amd              up   infinite     13   idle hermes[1-4],artemis[1-6],trillian[1-3]

You can also see what resources (such as GPUs) that a node has available by running the command scontrol show node <nodename>.

abc1de@portal01 ~$ scontrol show node ai10
NodeName=ai10 Arch=x86_64 CoresPerSocket=8
   CPUAlloc=0 CPUTot=32 CPULoad=0.01
   AvailableFeatures=(null)
   ActiveFeatures=(null)
   Gres=gpu:nvidia_geforce_gtx_1080:4
   NodeAddr=ai10 NodeHostName=ai10 Version=20.11.9
   OS=Linux 3.10.0-957.5.1.el7.x86_64 #1 SMP Fri Feb 1 14:54:57 UTC 2019
   RealMemory=128000 AllocMem=0 FreeMem=120054 Sockets=2 Boards=1
   State=IDLE ThreadsPerCore=2 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A
   Partitions=gnolim
   BootTime=2023-02-15T10:01:04 SlurmdStartTime=2023-03-14T12:29:33
   CfgTRES=cpu=32,mem=125G,billing=32
   AllocTRES=
   CapWatts=n/a
   CurrentWatts=0 AveWatts=0
   ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s
   Comment=(null)

To use SLURM resources, you must submit your jobs (program/script/etc.) to the SLURM controller. The controller will then send your job to compute nodes for execution, after which time your results will be returned. There is also an direct login option (see below) that doesn't require a job script.

Users can submit SLURM jobs from portal.cs.virginia.edu. You can submit jobs using the commands srun or sbatch. Let's look at a very simple example script and sbatch command.

Here is our script, all it does is print the hostname of the server running the script. We must add SBATCH options to our script to handle various SLURM options.

#!/bin/bash
# --- this job will be run on any available node
# and simply output the node's hostname to
# my_job.output
#SBATCH --job-name="Slurm Simple Test Job"
#SBATCH --error="my_job.err"
#SBATCH --output="my_job.output"
echo "$HOSTNAME"

We run the script with sbatch and the results will be put in the file we specified with –output. If no output file is specified, output will be saved to a file with the same name as the SLURM jobid.

[abc1de@portal04 ~]$ sbatch slurm.test
Submitted batch job 640768
[abc1de@portal04 ~]$ more my_job.output
cortado06

Here is a similar example using srun running on multiple nodes:

abc1de@portal01 ~ $ srun -w slurm[1-5] -N5 hostname
slurm4
slurm1
slurm2
slurm3
slurm5

If the node to be used is NOT in the main (default) “partition” (or queue), then you must specify the partition in your job script:

#!/bin/bash
# --- this job will be run on any available node in the "gpu" partition
# and simply output the node's hostname to
# my_job.output
#SBATCH --job-name="Slurm Simple Test Job"
#SBATCH --error="my_job.err"
#SBATCH --output="my_job.output"
# --- specify the partition (queue) name
#SBATCH --partition="gpu"
echo "$HOSTNAME"

If you are trying to use a node that is not in the default partition, and you don't specify the partition in your job script, you will get a message from srun saying “queued and waiting for resources”, but the job will not start.

You can use srun to login directly to a server controlled by the SLURM job scheduler. This can be useful for debugging purposes as well as running your applications without using a job script.

We must pass the –pty option to srun so output is directed to a pseudo-terminal.

For example, to open a direct login job on the node “cortado04”, use:

abc1de@portal ~$ srun -w cortado04 --pty bash -i -l -
abc1de@cortado04 ~$ hostname
cortado04
abc1de@cortado04 ~$

The -w argument selects the server into which to login. The -i argument tells bash to run as an interactive shell. The -l argument instructs bash that this is a login shell, this, along with the final - are important to reset environment variables that otherwise might cause issues using Environment Modules

If a node is in a partition (see below for partition information) other than the default “main” partition (for example, the “gpu” partition), then you must specify the partition in your command, for example:

abc1de@portal ~$ srun -w lynx05 -p gpu --pty bash -i -l -

If you are using a reservation, you must specify the –reservation=<reservationname> option to srun. Be sure to use -p <partition name> if the node you want to use is in gpu, nolim, gnolim or any other non-default partition.

Please be aware of jobs you start and make sure that they finish executing. If your job does not exit gracefully, it will continue running on the server, taking up resources and preventing others from running their jobs.

To cancel a running job, use the scancel [jobid] command

abc1de@portal01 ~ $ squeue
             JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
            467039      main    sleep    abc1de  R       0:06      1 artemis1           <--  Running job
abc1de@portal01 ~ $ scancel 467039

The default signal sent to a running job is SIGTERM (terminate). If you wish to send a different signal to the job's processes (for example, a SIGKILL which is often needed if a SIGTERM doesn't terminate the process), use the –signal argument to scancel, i.e.:

abc1de@portal01 ~ $ scancel --signal=KILL 467039

Slurm refers to job queues as partitions. We group similar systems into separate queues. For example, there is a “main” queue for general purpose systems, and a “gpu” queue for systems with GPUs. These queues can have unique constraints such as compute nodes, max runtime, resource limits, etc.

If no partition is specified in your job script or when using the 'srun' command, it will go to the default partition “main”.

The “main” and “gpu” partitions have a time limit set, so jobs will terminate after a specified number of days, as shown in the output of the 'sinfo' command. However, there are two additional partitions, “nolim” and “gnolim”, that have a time limit of 20 days, which is effectively unlimited time.

Partition is indicated by -p partname or –partition partname.

To specify a partition with sbatch file:

#SBATCH --partition=gpu

Or from the command line with srun

-p gpu

An example running the command hostname on the main partition, this will run on any node in the partition:

srun -p main hostname

If a job is expected to run longer than the default for a given partition, two other paritions with unlimited runtime named nolim and gnolim exist for jobs that have long runtimes.

The partition nolim is for long running jobs that do not require a GPU, and gnolim is for long running jobs that require a GPU.

To utilize these partitions, simply specify the name of the partition in srun or sbatch.

Slurm handles GPUs and other non-CPU computing resources using what are called GRES Resources (Generic Resource). To use the GPU(s) on a system using Slurm, either using sbatch or srun, you must request the GPUs using the –gres:x option. You must specify the gres flag followed by : and the quantity of resources

Say we want to use 4 GPUs on a system, we would use the following sbatch option:

#SBATCH --gres=gpu:4

Or from the command line

--gres=gpu:4

Reservations for specific resources or nodes can be made by submitting a request to cshelpdesk@virginia.edu. For more information about using reservations, see the main article on SLURM Reservations

The SLURM scheduler implements the Accounting features of slurm. So users can execute the sacct command to find job accounting information, like job ids, job names, partition run upon, allocated CPUs, job state, and exit codes. There are numerous other options supported. Type man sacct on portal to see all the options.

Due to the way sbatch spawns a bash session as a no-login session, initialization files are not loaded from /etc/profile.d. This prevents the initialization of the Environment Modules system and will prevent you from loading software modules.

To fix this, simply include the following line in your sbatch scripts:

source /etc/profile.d/modules.sh

The SLURM scheduler allows for preemption on certain nodes within the cluster. To view which nodes that have preemption enabled, execute the sinfo –format=“%P %N” | grep hipri. Any nodes within a partition with hipri for high priority in its name, have preemption enabled.

If you submit a job to a node in a high prioirty partition, it is possible it will be suspended if the owner of the node or their research group submits to it while your job is running.

This does not apply if a node is reserved.

To use a high priority partition, simply specify the partition name in your srun commands or sbatch scripts. This is done using the partition flag -p <partition name> or –partition=<partition name>.