Neocortex SDFlex

Neocortex SDFlex uses Slurm for interactive and batch workloads. The sdf partition contains the two SDF execution environments, sdf-1 and sdf-2. Use the login node only for light tasks such as editing files, inspecting allocations, and submitting or monitoring jobs.

Before submitting work, confirm the correct project and Unix group:

projects groups newgrp GROUPID

Common Slurm commands include:

squeue -u $USER # Show your queued and running jobs squeue -j JOB_ID # Show one job scancel JOB_ID # Cancel a job scontrol show job JOB_ID # Display detailed job information

For an interactive session on an available SDF node, use:

interact

To request a specific SDF execution environment:

srun --nodelist=sdf-1 --pty bash -i

A minimal CPU-only SDFlex batch job can use:

#!/bin/bash #SBATCH --job-name=sdf-example #SBATCH --partition=sdf #SBATCH --account=GROUPID #SBATCH --time=01:00:00 #SBATCH --cpus-per-task=28 #SBATCH --output=%x-%j.out newgrp GROUPID echo "Running on $(hostname)" echo "Local storage: $LOCAL" ./my_program

Submit the script and monitor it with:

sbatch job.sbatch squeue -u $USER

CPU-only preprocessing, compilation, data preparation, and memory-intensive SDFlex workloads should not request a Cerebras accelerator. A CS-2 training or evaluation workflow that requires a Cerebras system must additionally request the Cerebras generic resource:

#SBATCH --gres=cs:cerebras:1 #SBATCH --ntasks=7 #SBATCH --cpus-per-task=14

The default wall-time value documented for Neocortex reservations is four hours, and the maximum supported wall time is 48 hours. Request only the time needed because shorter jobs are generally easier to schedule.

See the instructions for running interactive, batch, validation, compilation, and Cerebras jobs on Neocortex, the complete Neocortex Slurm and Cerebras job-script examples, and the Neocortex reservation and wall-time instructions.