BeaverDeck logo
BeaverDeck

Home / BeaverDeck / Docs / Insights Guide / GPU Insights / GPU Fragmentation

GPU Fragmentation

BeaverDeck uses this check to identify a specific gpu condition that may need operator review.

Permissions: viewing checks requires insights: view. Opening a linked object or logs requires the corresponding resource permission, and the BeaverDeck ServiceAccount must be allowed to read the Kubernetes resources used by the check. Suppressing a finding requires insights: edit and affects all users.
Check typegpu-fragmentation
Insights sectionGPU Insights
Alert severityWarning

When It Reports A Finding

A GPU Pod has been Pending for at least 5 minutes, total free GPU capacity across schedulable GPU nodes is sufficient, but no single node has enough free GPUs for that Pod.

Why This Is A Problem

Capacity exists but is stranded across nodes, so a multi-GPU workload cannot schedule and additional GPU nodes may be added unnecessarily.

Recommended Response

  1. Compare the pod's GPU request with free capacity on each schedulable GPU node.
  2. Consolidate or reschedule existing GPU workloads where disruption policies allow.
  3. Reduce the request, split the workload, or add a node shape that can satisfy the per-pod requirement.

Scope And Limitations

The check uses GPU request accounting. Other scheduler predicates such as affinity, taints, topology, CPU, memory, and storage can also block the pod.

After remediation: refresh GPU Insights and verify the underlying resource or metric. Suppress the finding only when the condition is intentional and its risk is accepted.