LSF to Slurm quick reference

From ScientificComputing
Jump to: navigation, search

Introduction

The commands for Slurm are similar to the ones used in LSF. You can find a mapping of the relevant commands below.

Job submission

Simple command

LSF Slurm
bsub command sbatch --wrap=command
bsub "command1 ; command2" sbatch --wrap="command1 ; command2"
bsub "command1 | command2" sbatch --wrap="command1 | command2"
bsub [LSF options] command sbatch [slurm options] --wrap="command"

Frequently used bsub/sbatch options

Parameter bsub sbatch
Job name -J job_name -J job_name   or   --job-name=job_name
Job array consisting of N sub-jobs -J job_name[1-N]" -a 1-N   or   --array=1-N
Ouput file (stdout) -o file_name
(default: lsf.oJOBID)
-o file_name   or   --output=file_name
(default: slurm-JOBID.out)
Error file (stderr) -e file_name
(default: merged with output file)
-e file_name   or   --error=file_name
(default: merged with output file)
Wall-clock time (default: 4h) -W HH:MM -t DD-HH[:MM]   or   --time=MM   or   --time=HH:MM:SS
Number of cores (default: 1) -n cores -n cores   or   --ntasks=cores
Number of cores per node -R "span[ptile=cores_per_node]" --ntasks-per-node=cores_per_node
Memory per core (default: 1024 MB) -R "rusage[mem=MB]" --mem-per-cpu=MB   (can also be expressed in GB using "G" suffix)
Number of GPUs (default: 0) -R "rusage[ngpus_excl_p=N]" -G N   or   --gpus=N
Memory per GPU -R "select[gpu_mtotal0>=MB]" --gres=gpumem:MB   (can also be expressed in GB using "G" suffix)
Local scratch space per core -R "rusage[scratch=MB]" not available
Local scratch space per node not available --tmp=MB   (can also be expressed in GB using "G" suffix)
Run job under a specific shareholder group -G shareholder_group -A shareholder_group   or   --account=shareholder_group
Notify user by email when job starts -B --mail-type=BEGIN
Notify user by email when job ends -N --mail-type=END,FAIL
(multiple types can be combined in one option, e.g. --mail-type=BEGIN,END,FAIL)

Shell script

LSF Slurm
bsub [options] < jobscript.sh sbatch [options] < jobscript.sh   or  
sbatch [options] jobscript.sh [arguments]
Job parameters can be passed as options to bsub or placed inside jobscript.sh using #BSUB pragmas:
#!/bin/bash

#BSUB -n 4
#BSUB -W 08:00
#BSUB -R "rusage[mem=2000]"
#BSUB -R "rusage[scratch=1000]"    # per core
#BSUB -J analysis1
#BSUB -o analysis1.out
#BSUB -e analysis1.err

module load xyz/123
command1
command2
...
Job parameters can be passed as options to sbatch or placed inside jobscript.sh using #SBATCH pragmas:
#!/bin/bash

#SBATCH -n 4
#SBATCH --time=8:00
#SBATCH --mem-per-cpu=2000
#SBATCH --tmp=4000                        # per node!!
#SBATCH --job-name=analysis1
#SBATCH --output=analysis1.out
#SBATCH --error=analysis1.err

module load xyz/123
command1
command2
...

Note:

  • In LSF, the jobscript.sh must be passed to bsub via the "<" operator
  • In LSF, scratch space is expressed per core, while in Slurm it is per node
  • In LSF, the default output file is "lsf.oJOBID", while in Slurm it is "slurm-JOBID.out"

Interactive job

LSF Slurm
bsub -Is [LSF options] bash srun --pty bash

Parallel job

LSF Slurm
bsub -n 256 -R "span[ptile=128]" sbatch -n 256 --ntasks-per-node=128   or  
sbatch -n 256 --nodes=2

The Slurm options

  • --ntasks-per-core,
  • --cpus-per-task,
  • --nodes, and
  • --ntasks-per-node

are supported.

Please note that for larger parallel MPI jobs that use more than a single node (more than 128 cores), you should add the sbatch option

-C ib

to make sure that they get dispatched to nodes that have the infiniband highspeed interconnect, as this will result a much better performance.

Job array

LSF Slurm
bsub -J jobname[1-N]" sbatch --array=1-N
bsub -J jobname[1-N%step]" sbatch --array=1-N:step
Environment variables defined in each job:
  • index of current sub-job: LSB_JOBINDEX
  • maximum index (N): LSB_JOBINDEX_END
  • step: LSB_JOBINDEX_STEP
Environment variables defined in each job:
  • number of sub-jobs: SLURM_ARRAY_TASK_COUNT
  • index of current sub-job: SLURM_ARRAY_TASK_ID
  • minimum index: SLURM_ARRAY_TASK_MIN
  • maximum index: SLURM_ARRAY_TASK_MAX

LSF example:

bsub -J "myarray[1-4]" 'echo "Hello, I am task $LSB_JOBINDEX of $LSB_JOBINDEX_END"'

Slurm example:

sbatch --array=1-4 --wrap='echo "Hello, I am task $SLURM_ARRAY_TASK_ID of $SLURM_ARRAY_TASK_COUNT"'

GPU job

LSF Slurm
bsub -R "rusage[ngpus_excl_p=1]" sbatch --gpus=1

For multi-node jobs you need to use the --gpus-per-node option instead.

GPU job requiring a specific GPU model

LSF Slurm
bsub -R "rusage[ngpus_excl_p=1]" -R "select[gpu_model0==NVIDIAGeForceGTX1080]" sbatch --gpus=gtx_1080:1
bsub -R "rusage[ngpus_excl_p=1]" -R "select[gpu_model0==NVIDIAGeForceRTX3090]" sbatch --gpus=rtx_3090:1
  • For Slurm, currently the specifiers gtx_1080 and rtx_3090 are supported until we add more GPU types.

GPU job requiring a given amount of GPU memory

LSF Slurm
bsub -R "rusage[ngpus_excl_p=1]" -R "select[gpu_mtotal0>=20480]" sbatch --gpus=1 --gres=gpumem:20g

The default unit for gpumem is bytes. You are therefore advised to specify units, for example 20g or 11000m.

Submit a job using a specific share

LSF Slurm
bsub -G es_example sbatch -A es_example

In Slurm, one can define a default share using the command: "echo account=es_example >> $HOME/.slurm/defaults"

Submit a job on a specific CPU model

LSF Slurm
bsub -R "select[model==EPYC_7H12]" sbatch --constraint=EPYC_7H12

Job chains

LSF Slurm
bsub -J job_chain
bsub -J job_chain -w "done(job_chain)"
sbatch -J job_chain -d singleton

Job dependencies

LSF Slurm
Job #1: bsub -J job1 command1
Job #2: bsub -J job2 -w "done(job1)" command2
Job #1: myjobid=$(sbatch --parsable -J job1 --wrap="command1")
Job #2: sbatch -J job2 -d afterany:$myjobid --wrap="command2"

In Slurm, sbatch --parsable returns the JOBID of the job

Job control

Job status

LSF Slurm
bjobs [JOBID] squeue [-j JOBID]
bjobs -p squeue -u USERNAME -t PENDING
bjobs -r squeue -u USERNAME -t RUNNING

Resource usage

LSF Slurm
bbjobs [JOBID] myjobs -j JOBID
scontrol show jobid -dd JOBID
sacct -l -j JOBID for finished jobs
sstat [--all] JOBID for running jobs

Use --format JobID,User,State,AllocCPUS,Elapsed,NNodes,NTasks,TotalCPU,REQMEM,MaxRSS,ExitCode instead of -l for a customizable, more readable output.

Killing a job

LSF Slurm
bkill [JOBID] scancel [JOBID]

Environment variables

LSF Slurm
$LSB_JOBID $SLURM_JOB_ID
$LSB_SUBCWD $SLURM_SUBMIT_DIR