Billing & Metering
Business Value: Know exactly what each tenant, project, and user consumes — down to GPU-hours, storage bytes, and network bandwidth — with transparent, auditable billing records.
What Gets Metered
| Resource | What's Measured | How | Granularity |
|---|---|---|---|
| GPU Compute | GPU-hours, utilization %, memory, power draw, temperature | GPU telemetry exporter on every node | Per GPU, configurable intervals |
| CPU Compute | CPU-hours, utilization %, core count, memory | Node exporter on every node | Per node, configurable intervals |
| Bare Metal Allocation | Node-hours: allocation start → release | Orchestration DB timestamps | Per node, precise timestamps |
| K8s Cluster Runtime | Cluster-hours: creation → deletion | Orchestration DB lifecycle events | Per cluster |
| Slurm Jobs | Per-job GPU-hours, CPU-hours, elapsed time, memory, exit code | Slurm accounting daemon → relational database | Per job |
| Storage (Parallel FS) | Directory size, quota utilization, I/O throughput | Storage reporting + storage plugin | Per tenant directory |
| Storage (Object/Platform) | PVC usage, S3 bucket size, object count | Storage management + CSI metrics | Per PVC, per bucket |
| Network (InfiniBand) | Bandwidth per tenant, RDMA throughput, packet drops | Fabric manager telemetry + IB port counters | Per partition (per tenant) |
Metering Pipeline
Raw metrics flow through a multi-stage pipeline: hardware-level exporters collect data at configurable intervals → time-series database scrapes and stores short-term → long-term HA storage provides retention → metering service aggregates into billable records per tenant/project/user → billing reports generated and exportable as CSV/PDF.
Quota Management
Pre-set resource limits prevent overspend and ensure fair capacity distribution. Quotas are enforced in real-time — if a tenant exceeds their quota, new resource creation is blocked. The portal dashboard shows quota usage with color-coded warnings at configurable thresholds.