How to Request Interactive Sessions on Compute Nodes

Environment

HPC4 cluster

Superpod cluster

SLURM workload manager

Any resource-intensive task (compilation, testing, debugging)

Issue

Compiling GPU code requires access to GPU hardware not available on login nodes

Resource-intensive tasks (compilation, large file operations, testing) should not run on login nodes

Login nodes have minimal resource limit allocated per user

Users need an interactive shell on compute nodes for development and testing work

Direct access to compute node resources is required for building software or debugging

Resolution

Use the srun command with --pty bash to request an interactive session on a compute node.

Basic Interactive Session

For HPC4 cluster:

# Request interactive session on AMD partition
srun --account=exampleproj \
     --partition=amd \
     --nodes=1 \
     --ntasks-per-node=1 \
     --cpus-per-task=256 \
     --time=4:00:00 \
     --pty bash

# Or request session on Intel partition
srun --account=exampleproj \
     --partition=intel \
     --nodes=1 \
     --ntasks-per-node=1 \
     --cpus-per-task=128 \
     --time=4:00:00 \
     --pty bash

For Superpod cluster:

# Request interactive session with GPU
srun --account=exampleproj \
     --partition=gpu \
     --nodes=1 \
     --ntasks-per-node=1 \
     --cpus-per-task=32 \
     --gpus-per-task=1 \
     --time=4:00:00 \
     --pty bash

Note

Replace exampleproj with your actual project account. You can check your available accounts with sacctmgr show assoc user=$USER format=account%20.

Common Options

Option	Description
`--account=<project>`	Project account to charge for resource usage (required)
`--partition=<name>`	Partition to use (amd, intel, gpu, etc.)
`--nodes=<n>`	Number of nodes to allocate
`--ntasks-per-node=<n>`	Number of tasks (processes) per node
`--cpus-per-task=<n>`	Number of CPU cores per task
`--gpus-per-task=<n>`	Number of GPUs per task (GPU partitions only)
`--time=<HH:MM:SS>`	Maximum wall time for the session
`--pty bash`	Request a pseudo-terminal with bash shell

Important

Do not specify --mem or --mem-per-cpu options. Memory is automatically allocated proportionally based on the number of CPUs or GPUs requested.

Use Cases

Compiling Software

See Application Compile Notes for detailed compilation instructions of some common applications.

GPU Development and Testing

# Request GPU resources for testing
srun --account=exampleproj \
     --partition=gpu \
     --gpus-per-task=1 \
     --cpus-per-task=16 \
     --time=1:00:00 \
     --pty bash

nvidia-smi

Exiting Interactive Sessions

To exit an interactive session and return to the login node:

# Exit the interactive session
exit

Warning

Your session will automatically terminate when the time limit is reached. Reserve enough time for your work and save your work frequently.

Best Practices

Use the correct partition: Consider production CPU Model, GPU used when selecting development partition
Monitor resource usage: Use hostname, and nvidia-smi to verify you’re using allocated resources
Clean up after yourself: Exit sessions when done to stop billed resource usage
Use tmux/screen on login node: These tools allow you to get back to your interactive session if your connection drops

Root Cause

Login nodes are shared resources with limited CPU, memory, and I/O capacity. Running resource-intensive tasks on login nodes:

May trigger automatic process termination
Violates cluster usage policies
Degrades performance for all users
Can cause system instability

Compute nodes are designed for heavy computational workloads and provide:

Matched architecture for optimized builds
Access to specialized hardware (GPUs, high-core-count CPUs)
Dedicated resources for your tasks
Better performance for compilation and testing
Isolation from other users’ work

References

Example Scripts

SLURM Documentation