Quotas & Limits¶
DC Suite bounds consumption per tenant so one team can't exhaust the fleet. This page explains the limits you'll run into and how to work with them.
Tenant quotas¶
Each tenant has quotas set by an administrator:
| Quota | Meaning |
|---|---|
| Max nodes | The most nodes this tenant can hold across all its clusters at once. |
| Max clusters | The most concurrent clusters this tenant can have. |
Your own quota is reported to the console; via the API it's surfaced on your
identity (GET /v1/whoami) where the backend exposes it.
If a create or resize would exceed a quota, it's refused up front with a clear message — you're never left with a half-formed cluster because of a quota.
Fleet availability¶
Separate from your quota is physical availability: how many nodes of a profile are free in the pool right now. The create form shows this per profile. You can only reserve what's both within your quota and available in the fleet — whichever is smaller.
flowchart LR
A[Nodes you can request] --> B{min of}
B --> C[Your remaining quota]
B --> D[Profile availability now]
Interconnect-bound allocation units¶
Some profiles allocate in whole interconnect domains (for example a full NVLink domain) rather than single nodes, because splitting a domain would strand high-bandwidth links. For those profiles the console only offers valid counts, and resizes move a domain at a time.
Storage quota¶
Each cluster has a shared storage quota (GiB) you set at creation — the size of its shared filesystem. See Storage & Data.
Job-level limits¶
Within a cluster, the scheduler enforces its own limits — partition
time-limits, per-job GPU/node caps, and queue policies. A job requesting more
resources than the cluster has (or than a partition allows) sits pending. Use
sinfo and scontrol show partition to see the policy on your cluster.
Rate & request limits¶
The API applies sensible request limits to protect the control plane. Handle
429 Too Many Requests by backing off and retrying; automation should poll
operations at a modest interval (e.g. every 10s) rather than in a tight loop.
Raising a limit¶
Quotas are an administrator setting. If you consistently hit max_nodes or
max_clusters, ask an admin to raise the tenant's quota — and see
Cost & Billing so the spend implications are understood first.
Next: Running Jobs.