Image Removed

Overview

SLURM

The new HPC system, Puma, uses SLURM as a job scheduler rather than PBS Pro. SLURM has several advantages:

• It provides more robust support for a larger number of jobs in queue.
• It is used by national HPC groups (XSEDE and TACC) making it easier for users to scale out to those systems.
• It has more sustainable support.

Allocations and Job Partitions (Queues)

Using Puma with SLURM is similar to using ElGato and Ocelote with PBS. Users will still receive a monthly allocation of cpu hours associated with their PI's group which will be deducted when they run their jobs in standard. Users will also still be able to use windfall to run jobs without consuming their monthly allocations. As on Ocelote and Puma, jobs run using windfall will still be subject to preemption when resources are requested by higher-priority jobs.

To request a specific partition (standard, windfall, or high_priority), see Job Partition Requests below.

The process of finding, loading, and using software as modules will not change on the new system. Users will still be able to utilize the standard commands described in the Software section in our User Guide. However, in a departure from our previous systems, modules will not be available to load and utilize on the login nodes. To load, use, and test software for job submissions, users will need to request an interactive session. Interactive sessions may be requested by simply using the command "interactive". 

Interactive command

When you are on a login node, you can request an interactive session on a compute node.  This is useful for checking available modules, testing submission scripts, compiling software, and running programs directly from the command line. To get an interactive session, we have a built in command that will allow you to quickly and easily do so by simply entering:

interactive

Submitting this actually runs the following:

salloc --job-name=interactive --mem-per-cpu=4GB --nodes=1   --ntasks=1 --time=01:00:00 --account=windfall --partition=windfall

If you find that this session is insufficient, interactive has built-in customization flags. For example, if you want to get a session faster, add your PI's account name to use the standard partition:

interactive -a account_name

Are you using X11 forwarding?

interactive -a account_name -x

Full usage:

interactive [-x] [-N nodes] [-n ncpus per node] [-Q optional qos] [-t hh::mm:ss] [-a account to charge]

Each GPU node on Puma offers up to four GPUs that can be reserved for your job. To request a GPU, you will include the resource name using the --gres SLURM directive. For example, if you wanted to request an interactive session with one GPU, you could run

salloc --job-name=interactive --mem-per-cpu=4GB --nodes=1 --ntasks=1 --time=01:00:00 --account=windfall --partition=windfall --gres=gpu:1

In a batch script, you would include the number of GPUs as an SBATCH directive. For example:

#SBATCH --gres=gpu:1

In both cases above, the jobs are requesting 1 GPU. This number can be increased up to 4 depending on the number of GPUs you need for your workflow.

High Memory Nodes

Puma has two high memory nodes available with 3TB of RAM each. These nodes have a ratio of 32GB of RAM per CPU, so a job requesting N CPUs would be allocated N*32GB of RAM. To request one, you may either explicitly set --mem-per-cpu=32gb, or  --constraint=hi_mem in your job script. For example, the following directives:

#SBATCH --ntasks=1
#SBATCH --cpus-per-task=5
#SBATCH --mem-per-cpu=32gb

would run the job on one of the high memory nodes with 160GB of RAM. The following would request identical resources:

#SBATCH --ntasks=1
#SBATCH --cpus-per-task=5
#SBATCH --constraint=hi_mem

PBS → SLURM Rosetta Stone

In general, SLURM can translate and execute scripts written for PBS. This means that if you submit a PBS script written for Ocelote or ElGato on Puma (with the necessary resource request modifications), your script will likely run. However, there are a few caveats that should be noted:

• You will need to submit your job with the new SLURM command, e.g. sbatch instead of qsub
• There may be some PBS directives that do not directly translate to SLURM which cannot be interpreted
• The environment variables specific to PBS and SLURM are different. If your job relies on these, you will need to update them. Common examples are PBS_O_WORKDIR and PBS_ARRAY_INDEX

To get acquainted with the new scheduling system, refer to the following list of common PBS commands, directives, and environment variables and their SLURM counterparts.

#SBATCH --nodes=<N>

Select N nodes

#SBATCH --gres=gpu:1

Unlike PBS, SLURM offers ways to make your job's output filenames more customizable through the use of character replacements. A table is provided below as a guide with some examples. Variables may be used or combined as desired. Note: character replacements may also be used with other SBATCH directives such as error filename, input filename, and job name.

#SBATCH --array=1-2
#SBATCH -o %A.out
#SBATCH --open-mode=append

Single serial job submission

PBS Script

#!/bin/bash
#PBS -N Sample_PBS_Job
#PBS -l select=1:ncpus=1:mem=1gb
#PBS -l walltime=00:01:00
#PBS -q standard
#PBS -W group_list=<group_name>


cd $PBS_O_WORKDIR
pwd; hostname; date

module load python
python --version

SLURM Script

#!/bin/bash
#SBATCH --job-name=Sample_Slurm_Job
#SBATCH --ntasks=1              
#SBATCH --mem=1gb                     
#SBATCH --time=00:01:00    
#SBATCH --partition=standard
#SBATCH --account=<group_name>    

# SLURM Inherits your environment. cd $SLURM_SUBMIT_DIR not needed
pwd; hostname; date

module load python
python --version

Array Submission

IMPORTANT:

When submitting jobs with named output files (i.e. with the line #SBATCH -o=Job.out) as arrays, SLURM will write every array element to that filename leaving you with only the output of the last completed job in the array. Use one of the following SLURM directives in your script to prevent this behavior:

Differentiates output files using array indices. Similar to PBS default. See SLURM Output Filename Patterns above for more information.

#SBATCH --output=Job-%a.out 

Appends the output from all tasks in an array to the same output file. Warning: if a file exists with that name prior to running your job, the output will be appended to that file

#SBATCH --open-mode=append

PBS Script

#!/bin/bash
#PBS -N Sample_PBS_Job
#PBS -l select=1:ncpus=1:mem=1gb
#PBS -l walltime=00:01:00
#PBS -q standard
#PBS -W group_list=<group_name>
#PBS -J 1-5


cd $PBS_O_WORKDIR
pwd; hostname; date

echo "./sample_command input_file_${PBS_ARRAY_INDEX}.in"
 

SLURM Script

#!/bin/bash
#SBATCH --output=Sample_SLURM_Job-%a.out
#SBATCH --ntasks=1              
#SBATCH --mem=1gb                     
#SBATCH --time=00:01:00    
#SBATCH --partition=standard
#SBATCH --account=<group_name>    
#SBATCH --array 1-5

# SLURM Inherits your environment. cd $SLURM_SUBMIT_DIR not needed
pwd; hostname; date

echo "./sample_command input_file_${SLURM_ARRAY_TASK_ID}.in"

MPI Example

For openmpi the important variables are set by default, so you do not need to include them in your scripts.

export SBATCH_GET_USER_ENV=1
export OMPI_MCA_btl_openib_cpc_include=rdmacm
export OMPI_MCA_btl_openib_if_include=bnxt_re1
export OMPI_MCA_btl_openib_rroce_enable=1
export OMPI_MCA_btl=vader,self,openib
export OMPI_MCA_oob_tcp_if_include=eth1

For Intel MPI, these variables are set for you:

export I_MPI_FABRICS=shm:ofi
export FI_PROVIDER=verbs
export FI_VERBS_IFACE=eth1

PBS Script

#!/bin/bash
#PBS -N Sample_MPI_Job
#PBS -l select=1:ncpus=16:mem=16gb
#PBS -l walltime=00:10:00
#PBS -W group_list=<group_name>
#PBS -q standard






cd $PBS_O_WORKDIR
pwd; hostname; date
module load openmpi   
/usr/bin/time -o mpit_prog.timing mpirun -np 16 a.out

SLURM Script

#!/bin/bash
#SBATCH --job-name=Sample_MPI_Job
#SBATCH --ntasks=16
#SBATCH --ntasks-per-node=16
#SBATCH --nodes=1
#SBATCH --mem-per-cpu=1gb
#SBATCH --time=00:10:00
#SBATCH --account=<group_name>
#SBATCH --partition=standard
#SBATCH --output=Sample_MPI_Job_%A.out
#SBATCH --error=Sample_MPI_Job_%A.err

# SLURM Inherits your environment. cd $SLURM_SUBMIT_DIR not needed
pwd; hostname; date
module load openmpi3
/usr/bin/time -o mpit_prog.timing mpirun -np 16 a.out

(ocelote) [netid@junonia ~]$ interactive
Run "interactive -h for help customizing interactive use"
Submitting with /usr/local/bin/salloc --job-name=interactive --mem-per-cpu=4GB --nodes=1    --ntasks=1 --time=01:00:00 --account=windfall --partition=windfall
salloc: Pending job allocation 531843
salloc: job 531843 queued and waiting for resources
salloc: job 531843 has been allocated resources
salloc: Granted job allocation 531843
salloc: Waiting for resource configuration
salloc: Nodes i16n1 are ready for job
[netid@i16n1 ~]$ 

(ocelote) [netid@junonia ~]$ interactive -h
Usage: /usr/local/bin/interactive [-x] [-g] [-N nodes] [-m memory per core] [-n ncpus per node] [-Q optional qos] [-t hh::mm:ss] [-a account to charge]

export SBATCH_GET_USER_ENV=1
export OMPI_MCA_btl_openib_cpc_include=rdmacm
export OMPI_MCA_btl_openib_if_include=bnxt_re1
export OMPI_MCA_btl_openib_rroce_enable=1
export OMPI_MCA_btl=vader,self,openib
export OMPI_MCA_oob_tcp_if_include=eth1

module unload openmpi3 gnu8

srun -n $NPROCESSES --mpi=pmi2