akirakyle/SlurmClusterManager.jl

julia package for running code on slurm clusters

SlurmClusterManager.jl

This package provides support for using julia within the Slurm cluster environment. The code is adapted from ClusterManagers.jl with some modifications.

Usage

This script uses all resources from a Slurm allocation as julia workers and prints the id and hostname on each one.

#!/usr/bin/env julia

using Distributed, SlurmClusterManager
addprocs(SlurmManager())
@everywhere println("hello from $(myid()):$(gethostname())")

If the code is saved in script.jl it can be queued and executed on two nodes using 64 workers per node by running

sbatch -N 2 --ntasks-per-node=64 script.jl

Differences from ClusterManagers.jl

• Only supports Slurm (see this issue for some background).
• Requires that SlurmManager be created inside a Slurm allocation created by sbatch/salloc. Specifically SLURM_JOBID and SLURM_NTASKS must be defined in order to construct SlurmManager. This matches typical HPC workflows where resources are requested using sbatch and then used by the application code. In contrast ClusterManagers.jl will dynamically request resources when run outside of an existing Slurm allocation. I found that this was basically never what I wanted since this leaves the manager process running on a login node, and makes the script wait until resources are granted which is better handled by the actual Slurm queueing system.
• Does not take any Slurm arguments. All Slurm arguments are inherited from the external Slurm allocation created by sbatch/salloc.
• Output from workers is redirected to the manager process instead of requiring a separate output file for every task.