Kellogg Linux Cluster Reserve

Kellogg Linux Cluster Reserve#

Kellogg Linux Cluster Reserve (KLC Reserve) gives Kellogg researchers scheduled access to dedicated compute resources through the SLURM job scheduler — including GPU nodes and high-core-count batch nodes that are not available when you log directly into a standard KLC login node.

Reserve resources are accessed by submitting jobs to the kellogg SLURM partition from any KLC node. Unlike standard KLC login nodes (shared interactively by many users), Reserve resources are allocated exclusively to your job for its duration.

Standard KLC vs. KLC Reserve#

	Standard KLC login nodes	KLC Reserve
Access	SSH, OnDemand, VS Code, or FastX to a named node	Submit a SLURM job from any KLC node
Resources	Shared with all users on that node	Dedicated cores, memory, and GPUs for your job
Core limit	24 cores at normal priority per user	Request up to full node capacity in one job
GPUs	Not available on login nodes	H100, A100, and L40S GPU nodes
Connection	Must stay connected (or use `tmux`/FastX) for interactive work	Batch jobs run after you log out
Best for	Development, debugging, and logged production runs	Dedicated resources, GPUs, job arrays

For the decision framework, see When to Use KLC Reserve. For interactive work on login nodes, see the KLC User Guide.

Available Resources#

Resource type	Nodes	Key specs	Best for
GPU — H100	2	64 cores, 1 TB RAM, 4 × 80 GB GPUs per node	Large-scale LLM training/inference, deep learning
GPU — A100	1	64 cores, 2 TB RAM, 1 × 80 GB GPU per node	GPU-accelerated ML training and inference
GPU — L40S	1	64 cores, 2 × 48 GB GPUs per node	LLM inference, rendering, general GPU workloads
High-memory CPU	1	64 cores, 2 TB RAM	Very large in-memory datasets, parallel CPU jobs

Check the KLC Slurm Node Availability table on KLC Node Resources for current GPU and high-memory capacity. Contact rs@kellogg.northwestern.edu for capacity planning questions.

Getting Access#

KLC Reserve uses the same Northwestern NetID and KLC account as standard KLC login access. If you already have a KLC account, submit jobs to the kellogg partition without additional setup.

Note

Run groups after logging into any KLC node to see the SLURM allocation names you belong to. Use --account=kellogg in job scripts. Contact rs@kellogg.northwestern.edu if jobs fail to submit.

Submit Your First Job#

Log in to KLC via SSH, KLC OnDemand, or VS Code.
Write a job script with #SBATCH directives for account, partition, cores, memory, and wall time. See Submitting SLURM Jobs.
Submit the job:
```
sbatch myjob.sh
```
Monitor the job:
```
squeue -u $USER
```
Inspect output in the log file specified by #SBATCH --output.

For GPU jobs, add --gres=gpu:1 (or a specific card type). See GPU Jobs.

For large in-memory CPU jobs on the dedicated high-memory node, pin the job with --nodelist=qhimem0501. See High-Memory Jobs.

Documentation Map#

Goal	Page
Decide whether Reserve fits your workload	When to Use KLC Reserve
Write and submit CPU batch jobs	Submitting SLURM Jobs
Request and use GPU nodes	GPU Jobs
Run a large in-memory CPU job	High-Memory Jobs
Need more GPUs than Kellogg nodes provide	GPU Concepts and Options
Serve an open-source LLM with vLLM	vLLM Inference on KLC Reserve GPUs
Serve an open-source LLM with Ollama	Ollama Inference on KLC Reserve GPUs
Understand CPU vs. GPU concepts	GPU Concepts and Options
Interactive work on login nodes	When to Use KLC Reserve
Standard KLC overview	Kellogg Linux Cluster