使用 Slurm 和 OpenMPI

1. Slurm Usage

Slurm is a powerful workload manager and job scheduler for high-performance computing (HPC) clusters.

Job Submission

• Batch Job (sbatch): Submit a script for later execution.
• Interactive Job (srun): Run a job interactively in real-time.

# Submit a batch script
sbatch myscript.sh

# Start an interactive bash session
srun --pty /bin/bash

Resource Allocation

You can specify resources using command-line flags or #SBATCH directives in your script.

Resource	Flag	Example	Description
CPU	-c / --cpus-per-task	sbatch -c 4 script.sh	Request 4 CPUs per task.
Memory	--mem	sbatch --mem=8G script.sh	Request 8GB of memory per node.
GPU	--gres	sbatch --gres=gpu:2 script.sh	Request 2 GPUs.
Partition	-p / --partition	sbatch -p debug script.sh	Submit to the debug partition (queue).
Tasks	-n / --ntasks	sbatch -n 10 script.sh	Launch 10 tasks (processes).
Tasks/Node	--ntasks-per-node	sbatch --ntasks-per-node=5	Launch 5 tasks per node.

References

• Slurm Official Documentation
• Slurm Quick Start User Guide
• LLNL Slurm User Tutorial

2. OpenMPI Usage

OpenMPI is an open-source Message Passing Interface (MPI) implementation.

Basic Usage

To run an MPI program manually (outside of Slurm), you often need a hostfile to specify where to run processes.

Example hostfile:

node01 slots=4
node02 slots=4

Run Command:

mpirun -n 8 -hostfile hostfile ./your_app

Note

--allow-run-as-root is required if running as root (not recommended).

References

• OpenMPI Documentation
• MPI Tutorial

3. Running MPI Jobs with Slurm

When using Slurm, you typically do not need to manually specify nodes or hostfiles. Slurm integrates with MPI to handle process distribution automatically.

Example Batch Script

Create a script named job.sh:

#!/bin/bash
#SBATCH --job-name=mpi_job
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=16
#SBATCH --time=01:00:00
#SBATCH --partition=standard
#SBATCH --output=mpi_job_%j.out

# Load MPI module if needed
# module load openmpi

# Run the MPI program
# srun is preferred over mpirun within Slurm
srun ./my_mpi_program

Submission

sbatch job.sh

Slurm will allocate 2 nodes with 16 tasks each (32 tasks total) and launch my_mpi_program on them.