Login vs Execution Nodes

Contents

Login vs Execution Nodes#

Login Nodes vs Execution Nodes in SLURM: Where Commands Run on HPC Systems

Understanding where commands execute is fundamental when working on SLURM-based HPC systems such as Stampede3.

Confusion between login nodes and compute (execution) nodes is one of the most common sources of misunderstanding in HPC—and SLURM enforces this distinction strictly.

At a high level:

  • Login nodes are for interaction and orchestration

  • Compute nodes are for actual computation

  • SLURM always enforces this separation

This model applies to all SLURM jobs, regardless of how they are submitted.


Login Nodes

Where orchestration happens

Login nodes are intended for:

  • editing files

  • preparing job scripts

  • submitting jobs

  • lightweight inspection and setup

They are not intended for heavy computation.

Key points:

  • sbatch runs on the login node

  • no SLURM jobs execute on login nodes

  • CPU- or memory-intensive work is prohibited

Login nodes exist to protect the stability of the shared system.


Compute (Execution) Nodes

Where computation happens

All SLURM jobs execute on compute nodes:

  • batch scripts start here

  • applications run here

  • MPI ranks and threads run here

When a job is scheduled:

  • SLURM allocates compute nodes

  • the batch script is launched on one node in the allocation

  • all actual computation occurs on compute nodes only

Login nodes are never involved once execution begins.


What sbatch Actually Does

Submission vs execution

A common misconception is that sbatch runs a job.

In reality:

  • sbatch submits a job to SLURM

  • SLURM queues the job

  • when resources become available, SLURM launches the batch script on compute nodes

Important clarification:

sbatch always runs on a login node. Batch scripts never do.

This distinction is fundamental to how SLURM operates.


Key Takeaways#

  • Login nodes orchestrate

  • Compute nodes execute

  • sbatch runs on login nodes

  • Batch scripts and applications run on compute nodes

  • This model is always enforced by SLURM

Once this mental model is clear, many HPC behaviors become immediately predictable.