Glossary: Projects and tutorials#
This page defines terms used in Project settings and the tutorials Projects and quotas, Resource sharing and pre-emption, Monitoring workloads, and Resource utilization. It is scoped to those topics—not a full AMD Resource Manager dictionary.
Borrowed resources#
Resources your project is using that count against another project’s quota because that project was not using them at the time. When the owning project needs those resources back, workloads using borrowed resources may be pre-empted. See also Resource sharing.
Checkpointing#
Saving work in progress so a long-running job can resume after a pre-emption or restart instead of starting over. For workload checkpointing in AMD Resource Manager, see Workload checkpointing in Workload priority classes.
Claim (resources)#
When a workload obtains or reserves cluster resources (GPUs, CPU, memory, etc.) for its pods—for example after scheduling or when a higher-priority workload pre-empts a lower-priority one to take those resources. Related terms: Reclaim, Pre-emption. See Workload priority classes.
Guarantee (guaranteed quota)#
The minimum resources a project is entitled to—the same idea as Quota: a floor, not a ceiling. In the UI and docs you may see guaranteed minimum, Guaranteed quota tab, or ensured quota / ensured amount of resources—they refer to this guarantee. Workloads can still use more than that minimum when spare capacity exists (resource sharing).
Pre-emption (preemption)#
The system suspends workloads that are using borrowed resources when the project that owns the quota needs them back. When capacity becomes available again, those workloads can resume automatically. Spelling pre-emption is also used in the UI and some docs.
Priority class preemption#
Preemption driven by workload priority (low, medium, high per workload), distinct from quota-based preemption. See Workload priority classes.
Project#
A container for your work on the platform: projects are isolated—resources, workloads, and secrets in one project are not accessible from another. Each project has settings (quotas, storages, secrets, users) and an associated Kubernetes namespace.
Project settings#
The UI area where you manage quotas, storages, secrets, users, and general options (including deletion) for a project. See Project settings.
Quota#
A minimum set of resources guaranteed to a project—a floor, not a cap. There is no maximum: projects can use more than their quota when spare capacity exists. AMD Resource Manager lets you set GPU, CPU, memory, and disk quotas per project. New projects start with zero quota until you configure them.
Quota-based preemption#
When the cluster is full, workloads from projects over quota (using extra capacity through idle resources) may be pre-empted so projects within their quota can run. In-quota projects reclaim borrowed resources from those that exceeded their quota.
Reclaim#
When the project that owns the quota takes borrowed resources back from another project—often described as in-quota projects reclaiming resources from projects that were over quota, or GPUs being reclaimed in a scheduling scenario (for example when a higher-priority project needs capacity that was temporarily used elsewhere). Closely related to Quota-based preemption and Pre-emption.
Resource sharing#
Other projects can use idle resources that are part of another project’s quota when that quota is not fully utilized. That use is borrowed until the owning project needs it back.
Resource utilization#
How much of the cluster’s resources (CPU, GPU, memory, etc.) are in use—often viewed on the Project details or monitoring views in the tutorials.
Workload#
A running unit of work on the cluster (for example a deployment, job, or AIM). Workloads are subject to project quotas and priority settings regardless of how they are submitted (UI, kubectl, etc.).