Interactive Sessions¶

This section provides official guidance on initiating interactive sessions on the REPACSS high-performance computing system. Interactive sessions allow users to request computational resources in real time and execute commands directly on compute nodes. This method is particularly suitable for software testing, debugging, exploratory tasks, and graphical application workflows.

Warning

Interactive sessions should be used exclusively for development and debugging purposes. Once your job is ready for full execution, please submit it using a SLURM batch job and exit from the resources.

Requesting Interactive Resources¶

To start an interactive session, use the interactive command. This command wraps around the resource request process, performs additional setup, and launches a shell on a compute node once resources are allocated.

interactive -c 8 -p h100

This method is the recommended way to start interactive sessions on REPACSS.

Tip

Do not call salloc directly unless instructed otherwise. The interactive command performs important environment configuration steps that salloc alone does not handle.

Interactive Sessions on GPU Nodes¶

When working with graphical processing units (GPUs), use the interactive command with the appropriate GPU partition and core count. For example:

interactive -c 8 -p h100

Within the session, use srun to launch your GPU applications:

srun --gres=gpu:nvidia_h100_nvl:1 ./my_gpu_program

Warning

If GPU resources are not explicitly requested, applications relying on GPU acceleration (e.g., CUDA) may fail with an error such as:

no CUDA-capable device is detected

Interactive Sessions on CPU-Only Nodes¶

For CPU-only workloads, use the interactive command with a CPU-only partition such as zen4:

interactive -c 8 -p zen4

Once the session begins, you can run programs directly or use srun for launching parallel tasks:

srun ./my_cpu_program

For MPI applications, load the required MPI module and use srun with the appropriate number of tasks. For example, to run LAMMPS (a molecular dynamics simulator) with 4 MPI processes:

module load mpich/4.1.2
srun -n 4 lmp -in input.lmp

See MPI Implementations and Usage Guidance for more details on available MPI implementations and their usage.

Resource Allocation Timeout and Immediate Scheduling¶

By default, interactive job requests will time out if resource allocation is not completed within six (6) minutes. If needed, a custom wait limit can be applied using:

interactive -c 8 -p zen4 --immediate=600

This example waits up to 600 seconds (10 minutes) for compute resources.

Note

If your connection is lost, the interactive session will terminate and any unsaved progress may be lost.

Common Issues and Resolutions¶

GPU Execution Errors: Confirm that GPU resources are requested in the interactive session and in any srun commands.
Invalid Account Settings: Contact REPACSS Support.
Unavailable Resources: If nodes are unavailable, consider lowering resource requests or choosing a less busy partition.