Slurm Executor#

This executor plugin interfaces Covalent with HPC systems managed by Slurm. For workflows to be deployable, users must have SSH access to the Slurm login node, writable storage space on the remote filesystem, and permissions to submit jobs to Slurm.

To use this plugin with Covalent, simply install it using pip:

pip install covalent-slurm-plugin

The following shows an example of a Covalent configuration that is modified to support Slurm:

username = "user"
address = ""
ssh_key_file = "/home/user/.ssh/id_rsa"
remote_workdir = "/scratch/user"
cache_dir = "/tmp/covalent"

nodes = 1
ntasks = 4
cpus-per-task = 8
constraint = "gpu"
gpus = 4
qos = "regular"

cpu_bind = "cores"
gpus = 4
gpu-bind = "single:1"

The first stanza describes default connection parameters for a user who can connect to the Slurm login node using, for example:

ssh -i /home/user/.ssh/id_rsa

The second and third stanzas describe default parameters for #SBATCH directives and default parameters passed directly to srun, respectively.

This example generates a script containing the following preamble:

#SBATCH --nodes=1
#SBATCH --ntasks=4
#SBATCH --cpus-per-task=8
#SBATCH --constraint=gpu
#SBATCH --gpus=4
#SBATCH --qos=regular

and subsequent workflow submission with:

srun --cpu_bind=cores --gpus=4 --gpu-bind=single:1

To use the configuration settings, an electron’s executor must be specified with a string argument, in this case:

import covalent as ct

def my_task(x, y):
    return x + y

Alternatively, passing a SlurmExecutor instance enables custom behavior scoped to specific tasks. Here, the executor’s prerun_commands and postrun_commands parameters can be used to list shell commands to be executed before and after submitting the workflow. These may include any additional srun commands apart from workflow submission. Commands can also be nested inside the submission call to srun by using the srun_append parameter.

More complex jobs can be crafted by using these optional parameters. For example, the instance below runs a job that accesses CPU and GPU resources on a single node, while profiling GPU usage via nsys and issuing complementary commands that pause/resume the central hardware counter.

executor = ct.executor.SlurmExecutor(
        "qos": "regular",
        "time": "01:30:00",
        "nodes": 1,
        "constraint": "gpu",
        "module load package/1.2.3",
        "srun --ntasks-per-node 1 dcgmi profile --pause"
        "n": 4,
        "c": 8,
        "cpu-bind": "cores",
        "G": 4,
        "gpu-bind": "single:1"
    srun_append="nsys profile --stats=true -t cuda --gpu-metrics-device=all",
        "srun --ntasks-per-node 1 dcgmi profile --resume",

def my_custom_task(x, y):
    return x + y

Here the corresponding submit script contains the following commands:

module load package/1.2.3
srun --ntasks-per-node 1 dcgmi profile --pause

srun -n 4 -c 8 --cpu-bind=cores -G 4 --gpu-bind=single:1 \
nsys profile --stats=true -t cuda --gpu-metrics-device=all \
python /scratch/user/experiment1/

srun --ntasks-per-node 1 dcgmi profile --resume