Documentation

Install Sonar on your GKE cluster, then use the support page if you need help from Pebble.

What Sonar deploys

Sonar discovers GPU inference workloads (vLLM, SGLang, generic GPU pods) on Kubernetes and, when a workload opts in, runs a power-profiler Job that measures throughput versus GPU power.

ComponentPurpose
WatcherDiscovers GPU workloads and schedules profiler Jobs when opted in
PrometheusScrapes DCGM (NVIDIA) or AMD device metrics exporter
Adminer UIBrowse SQLite discovery and profiler results
Profiler JobsMeasure throughput at different GPU power levels (opt-in only)

Ready to deploy on GKE?

Start with the install guide — NVIDIA is the Marketplace default; AMD clusters use a Helm values overlay. Marketplace deploys use the app name and namespace you choose.