Expanse

When researchers submit a job, they have to say how much memory, CPU, GPU and time it needs. Most ask for more than they use so the job isn’t killed partway through, and the scheduler reserves all of it. The cluster looks full and the queue keeps growing, while much of what was reserved is never used. Expanse works with the scheduler you already run. It learns from your own jobs, on your own hardware, and works out what each job will actually need before it runs. That answer reaches your cluster in one of two ways. A researcher can ask for it before submitting and choose whether to use it. Or your scheduler asks Expanse as each job is submitted and applies the answer under rules your platform team sets, so researchers don’t size jobs at all. In both cases, researchers submit jobs the same way they do now. When a job fails, Expanse tells you why and shows you the evidence.

​

What you get

• Requests sized to the job before it runs, so fewer jobs die out of memory and less capacity sits reserved but idle.
• A clear reason when a job fails, with the evidence behind it and a prompt your engineers can hand to their coding agent.
• Requested against used capacity for every team and user, measured against your own scheduler’s accounting.

​

Where it runs

Expanse runs inside your network, on premises or in your cloud account. A collector runs on each node of your cluster, on Kubernetes and on SLURM alike. The models, your workload data and your source code stay on your hardware. What reaches Expanse is a short, documented list of calls about the health of the deployment and aggregate results, and an air-gapped deployment sends nothing at all. Security has the detail. We work with teams across many industries, with very different requirements, and we shape each deployment around your security rules.

​

Talk to us

The quickest way to find out whether Expanse fits your cluster is a short call. We’ll ask how you schedule today, show you what Expanse would see, and send you the deployment and security documentation afterwards.