slurm is albedo's job scheduling system. It is used to submit jobs from the login nodes to the compute nodes.
Jobs
Submitting jobs
- To work interactively on a compute node use salloc.
You can use all options (more CPU, RAM, time, partition, qos, ...) described in the next section.
To enable working with graphical interfaces (X forwarding) add the option --x11 . - Job scripts are submitted via sbatch
Specifying job resources
Job resources are defined at the header of your job script (or as command line arguments for sbatch or salloc). A full list see https://slurm.schedmd.com/sbatch.html#SECTION_OPTIONS. Here is a list of the most common ones:
#SBATCH --account=<account> # Your account
#SBATCH --partition=<partition> # Slurm Partition; Default: smp
#SBATCH --time=<time> # time limit for job; Default: 0:30:00
#SBATCH --qos=<QOS> # Slurm QOS; Default: 30min
#SBATCH --nodes=<#Nodes> # Number of nodes
#SBATCH --ntasks=<#Tasks> # Number of tasks (MPI) tasks to be launched
#SBATCH --mem=<memory> # If more than the default memory is needed;
# Default: <#Cores> * <mem per node>/<cores per node>
#SBATCH --ntasks-per-node=<ntasks> # Numer of tasks per node
#SBATCH --mail-user=<email adress> # Your mail adress if you want to get notifications
#SBATCH --mail-type=<email type> # Valid type values are NONE, BEGIN, END, FAIL, REQUEUE, ALL
#SBATCH --job-name=<jobname> # Job name
#SBATCH --output=<filename_pattern> # File where the standard output is written to(*)
#SBATCH --error=<filename_pattern> # File where the error messages are written to(*)
*) For filename patterns see: https://slurm.schedmd.com/sbatch.html#SECTION_%3CB%3Efilename-pattern%3C/B%3E
Job enforcements
We implemented some enforcements to improve albedo's overall performance.
- Jobs requesting --partition=fat but only low memory are rejected.
- Jobs requesting less than 40 nodes are enforced to use only nodes connected to the very same Infiniband switch (if this is feasible within 15 Minutes).
Details about specific parameters
Account (-A)
Compute resources are attributed to (primary)sections and projects at AWI. Therefore it is mandatory to specify an account.
This is new on Albedo, compared to ollie
The slurm accounts you may use are listed after login, with info.sh -s or can be shown via
sacctmgr -s show user name=$USER format=user,account%-30
Note: The account noaccount is just a dummy account that can not be used for computing.
You can change the default setting on your own:
sacctmgr modify user $USER set DefaultAccount=<account>
Partitions (-p)
Identical compute nodes are combined in partitions. More information about the hardware specification of each node can be found in the System Overview.
Partition | Nodes | Description |
|---|---|---|
smp | prod-[001-200] |
|
mpp | prod-[001-200] |
|
fat | fat-00[1-2] |
|
matlab | fat-00[3-4] | currently reserved for matlab users (as personal matlab licenses are node-bound). This might change later. To prohibit single users from allocating too many resources on these dedicated nodes, we limit the resources per user in this partition to 32 CPUs and 1TB RAM. Please get in touch with us if these limitations conflict with your use case! |
gpu | gpu-00[1-2] |
|
Quality of service (--qos)
Slurm's QOS is a way for us to influence a job's priority (priority QOS_factor) and "cost" (UsageFactor) based on the job's size (we only take walltime into account here!). We therefore created the different QOS, which are listed below.
The default QOS is 30min; for a job with a walltime >30min you have to select and set an appropriate QOS in addition to your walltime!
To facilitate development and testing, we have reserved 10 nodes during working hours exclusively for jobs with QOS=30min.
QOS | max. walltime | max. Nodes/User | UsageFactor | Priority QOS_factor | Notes |
|---|---|---|---|---|---|
30min | 00:30 | - | 1 | 1 | default |
12h | 12:00 | 120 | 1 | 0 | |
48h | 48:00 | 80 | 2 | 0 | |
1wk | 7-00:00:00 (168h) | 1 | 10 | 0 | only available for users upon request; whenever possible try to adapt your workflow to allow for shorter walltime! In case of urgent system maintenance we might cancel long jobs using this QOS without further warning! |
Job Scheduling
Priority
Jobs on albedo are scheduled based on a priority that is computed by Slurm depending on multiple factors (https://slurm.schedmd.com/priority_multifactor.html).
The higher the priority, the sooner your job begins. (In principle – the backfill scheduling plugin helps making best use of available resources by filling up resources that are reserved (and thus idle) for large higher priority jobs with small (lower priority) jobs.)
At AWI, only few of the possible factors are taken into account:
Job_priority = (PriorityWeightAge) * (age_factor) + (PriorityWeightFairshare) * (fair-share_factor) + (PriorityWeightQOS) * (QOS_factor) - nice_factor
The weights in this formula are set to balance the different factors and might become subject for tuning.
The current values can be assessed by running
$ scontrol show config | grep -i PriorityWeight PriorityWeightAge = 3500 PriorityWeightAssoc = 0 PriorityWeightFairShare = 10000 PriorityWeightJobSize = 0 PriorityWeightPartition = 0 PriorityWeightQOS = 5000 PriorityWeightTRES = (null)
The factors (except of the nice_factor (default is zero), which can be set by the user to downgrade the jobs priority by the setting --nice=...), are numbers in the range from 0 to 1.
They are shortly explained in the following.
You can check the recent usage of albedo with this command:
sreport -t Percent cluster UserUtilizationByAccount Start=$(date +%FT%T -d "1 week ago") Format=used,login,account FairShare
The fairshare factor is the most important factor here, but also the most difficult factor to understand. This factor is calculated using the "classic" fairshare algorithm of Slurm (https://slurm.schedmd.com/classic_fair_share.html). It computes the fairshare for each user based on the recent usage of the system.
Note, the usage of your associated account is *not* taken into accunt here, as it was the case on ollie!
Usage is basically "CPU seconds", but weighted using the UsageFactor depending on the used QOS (see section QOS). Furthermore, the usage taken into account here decays with time (with a half life time of 7 days).
Fairshare is the calculated by
FS = 2^(- (U_N / S_N) / D),
where the normalized usage U_N is the own usage relative to the total usage of albedo, the normalized share S_N is the share of a user on the entire system (1/(number of albedo users)) and D is a dampening factor. The formula basically assigns users a fairshare > 0,5 who under-use their share and < 0,5 for users who over-use their share. This is shown in the following figure, where the dots are taken from historic data from ollie. D has to be adjusted to account for the many users with an HPC account, who don't use it. This might also need tuning.
Fairshare values can be shown with the command
sshare
QOS
To reward usage of the short 30min QOS for jobs, which are easier to schedule, the priority is increased!
See section about QOS.
Age
Job's priority slowly increases with waiting time in the queue. With the current setting the priority is increased by 500 for each day waiting. The factor saturates after 7 days.
Note: Jobs waiting for a dependency to finish are not ageing.
Useful Slurm commands
- sinfo shows existing queues
- scontrol show job <JobID> shows information about specific job
- sstat <JobID> shows resources used by a specific job
- squeue shows information about queues and used nodes
- sbatch <script> submits a batch job
- salloc <resources> requests access to compute nodes for interactive use
- scancel <JobID> cancels a batch job
- srun <ressources> <executable> starts a (parallel) code
- sshare and sprio give information on fair share value and job priority
- sreport -t Percent cluster UserUtilizationByAccount Start=$(date +%FT%T -d "1 month ago") Format=used,login,account | head -20 top usage users during the last month
Do's & Don'ts
- Do not use srun for simple non-parallel jobs like cp, ln, rm, cat, g[un]zip
- Do not write loops in your slurm script to start several instance of similar jobs → See Job arrays below
- Make use of parallel srun p[gu]igz instead of g[un]zip if you have allocated more than one CPU already
- Do not allocate costly resources (like fat/gpu nodes) if you not need them. Check the CPU/Memory-Efficiency of your jobs with info.sh -S
Example Scripts
Job arrays
Job arrays in Slurm are an easy way to submit multiple similar jobs (e.g. executing the same script with multiple input data). See here for further details.
#!/bin/bash
#SBATCH --account=<account> # Your account
#SBATCH --partition=smp
#SBATCH --time=0:10:00
#SBATCH --ntasks=1
# run 100 tasks, but only run 10 at a time
#SBATCH --array=1-100%10
#SBATCH --output=result_%A_%a.out # gives result_<jobID>_<taskID>.out
echo "SLURM_JOBID: $SLURM_JOBID"
echo "SLURM_ARRAY_TASK_ID: $SLURM_ARRAY_TASK_ID"
echo "SLURM_ARRAY_JOB_ID: $SLURM_ARRAY_JOB_ID"
# Here we "translate" the $SLURM_ARRAY_TASK_ID (which takes values from 1-100)
# into an input file, that we want to analyze.
# Suppose 'input_files.txt' is a text file that has 100 lines, each containing
# the respective input file.
INPUT_LIST=input_files.txt
# Read the (SLURM_ARRAY_TASK_ID)th input file
INPUT_FILE=`sed -n "${SLURM_ARRAY_TASK_ID}p" < ${INPUT_LIST}`
srun my_executable $INPUT_FILE
How you “translate” your task ID into the srun command line is up to you. You could, for example, also have different scripts that you select in some way and execute.
MPI
#!/bin/bash #SBATCH --account=<account> # Your account #SBATCH --time=0:10:00 #SBATCH -p mpp #SBATCH -N 2 #SBATCH --tasks-per-node=128 #SBATCH --cpus-per-task=1 #SBATCH --hint=nomultithread #SBATCH --job-name=mpi #SBATCH --output=out_%x.%j # disable hyperthreading #SBATCH --hint=nomultithread module purge module load xthi/1.0-intel-oneapi-mpi2021.6.0-oneapi2022.1.0 intel-oneapi-mpi # module load xthi/1.0-openmpi4.1.3-gcc8.5.0 openmpi/4.1.3 ## Uncomment the following line to enlarge the stacksize if needed, ## e.g., if your code crashes with a spurious segmentation fault. # ulimit -s unlimited # To be on the safe side, we emphasize that it is pure MPI, no OpenMP threads export OMP_NUM_THREADS=1 srun xthi | sort -g -k 4
#!/bin/bash #SBATCH --account=<account> # Your account #SBATCH --time=0:10:00 #SBATCH -p mpp #SBATCH -N 2 #SBATCH --tasks-per-node=31 #SBATCH --hint=nomultithread #SBATCH --job-name=mpi_partial_node #SBATCH --output=out_%x.%j # disable hyperthreading #SBATCH --hint=nomultithread module purge module load xthi/1.0-intel-oneapi-mpi2021.6.0-oneapi2022.1.0 intel-oneapi-mpi # module load xthi/1.0-openmpi4.1.3-gcc8.5.0 openmpi/4.1.3 ## Uncomment the following line to enlarge the stacksize if needed, ## e.g., if your code crashes with a spurious segmentation fault. # ulimit -s unlimited # To be on the safe side, we emphasize that it is pure MPI, no OpenMP threads export OMP_NUM_THREADS=1 # The --cpu-bind=rank_ldom distributes the tasks via the node's cores # respecting the node's NUMA domains srun --cpu-bind=rank_ldom xthi | sort -g -k 4
OpenMP
#!/bin/bash #SBATCH --account=<account> # Your account #SBATCH --time=0:10:00 #SBATCH -p smp #SBATCH --tasks-per-node=1 #SBATCH --cpus-per-task=64 #SBATCH --job-name=openMP #SBATCH --output=out_%x.%j # disable hyperthreading #SBATCH --hint=nomultithread module purge module load xthi/1.0-intel-oneapi-mpi2021.6.0-oneapi2022.1.0 intel-oneapi-mpi # module load xthi/1.0-openmpi4.1.3-gcc8.5.0 openmpi/4.1.3 ## Uncomment the following line to enlarge the stacksize if needed, ## e.g., if your code crashes with a spurious segmentation fault. # ulimit -s unlimited # export OMP_STACKSIZE=128M # This binds each thread to one core export OMP_PROC_BIND=TRUE # OpenMP and srun, both need to know the number of CPUs per task export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK export SRUN_CPUS_PER_TASK=$SLURM_CPUS_PER_TASK srun xthi | sort -g -k 4
Hybrid (MPI+OpenMP)
#!/bin/bash #SBATCH --account=<account> # Your account #SBATCH --time=0:10:00 #SBATCH -p mpp #SBATCH -N 2 #SBATCH --tasks-per-node=8 #SBATCH --cpus-per-task=16 #SBATCH --job-name=hybrid #SBATCH --output=out_%x.%j # disable hyperthreading #SBATCH --hint=nomultithread module purge module load xthi/1.0-intel-oneapi-mpi2021.6.0-oneapi2022.1.0 intel-oneapi-mpi # module load xthi/1.0-openmpi4.1.3-gcc8.5.0 openmpi/4.1.3 ## Uncomment the following line to enlarge the stacksize if needed, ## e.g., if your code crashes with a spurious segmentation fault. # ulimit -s unlimited # export OMP_STACKSIZE=128M # This binds each thread to one core export OMP_PROC_BIND=TRUE # OpenMP and srun, both need to know the number of CPUs per task export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK export SRUN_CPUS_PER_TASK=$SLURM_CPUS_PER_TASK srun xthi | sort -g -k 4
Usage of GPU
#!/bin/bash #SBATCH --account=<account> # Your account #SBATCH --time=0:10:00 #SBATCH -p gpu #SBATCH --ntasks=1 #SBATCH --gpus=a100:2 # allocate 2 (out of 4) A100 GPUs; to get 2 (out of 2) A40 GPUs use --gpus=a40:2 #SBATCH --hint=nomultithread #SBATCH --job-name=gpu #SBATCH --output=out_%x.%j # disable hyperthreading #SBATCH --hint=nomultithread ## Uncomment the following line to enlarge the stacksize if needed, ## e.g., if your code crashes with a spurious segmentation fault. # ulimit -s unlimited # To be on the safe side, we emphasize that it is pure MPI, no OpenMP threads export OMP_NUM_THREADS=1 srun your_code_that_runs_on_GPUs