VEP 254: Guest GPU Metrics via VSOCK

[machadovilaca]

Primary contact (assignee):

/assign @machadovilaca

Current Feature Stage: New

Feature Gate: GPUMetrics

Responsible SIGs:

Primary SIG:
/sig observability

Additional SIGs (optional):
/sig compute

Enhancement link:

Timeline:

• Alpha release target (1.9):

• Beta release target (x.):

• Stable release target (x.y):

• Alpha

  • VEP PRs: VEP #254: Guest GPU Metrics via VSOCK #255
  • Code PRs:
  • Docs PRs:

Additional context:

Collect GPU metrics (utilization, memory, temperature, power, ECC errors) from inside guest VMs via a virtio-serial channel and expose them as kubevirt_vmi_gpu_* Prometheus metrics from virt-handler. A lightweight guest agent (gpu-metrics-agent) uses NVML to collect metrics and communicates with the host over a newline-delimited JSON protocol. Supports both Linux and Windows guests. See the VEP for full design details.

Important

Please keep this description up to date.

[enp0s3]

@machadovilaca Can you please update the issue title to align with the VEP title?