What you get
- Token layer: serving-stack token counters become tokens per second per server and tokens per kWh as colour across the room: which node produced the most tokens, and how many tokens any scenario would lose.
- Device layer: GPU utilisation, temperature, power and memory; node CPU and memory per host, shown on the server in its rack.
- Power and Air layers: IPMI watts and inlet temperatures; UPS runtime, load, battery and on-battery state on the one-line diagram. Click the UPS, press Fail, see what goes dark; minutes to 27, 32 and 35 °C are estimated.
What you can do today
In Grafana, open any panel and use Inspect → Data → Download CSV. You get time and value columns; add item (the device name) and metric columns and paste the rows into the Readings sheet. Metric names such as power_w, inlet_c, util_pct or ups_runtime_min drive the simulation directly; anything else is kept as a series you can bind by hand.
What the collector reads
| Exporter family | DC Sim metric | Layer |
|---|---|---|
| GPU telemetry exporter: utilisation, temperature, power, memory | GPU util %, GPU °C, GPU W, GPU memory | Device, Token |
| vLLM and SGLang: generation and prompt token counters | generated and prompt tokens (counter to rate) | Token |
| LiteLLM: total, input and output tokens, requests | tokens per model, requests | Token |
| node_exporter: CPU seconds, memory, hardware temperatures | CPU %, memory %, CPU °C | Device, Air |
| ipmi_exporter: power, temperatures (inlet, intake, ambient), fans | power W, inlet °C, fan rpm | Power, Air, Device |
| snmp_exporter with the standard UPS-MIB: minutes remaining, load, charge, output source | UPS runtime, load %, battery %, on battery | Power |
| any family matching your include patterns | raw series (counters become rates) to bind by hand | Import |
Later
| Item | Status |
|---|---|
| Live collector scraping your exporters through the agent | Roadmap (built, in testing) |
| Read from the Prometheus server itself, one target instead of one per exporter | Roadmap |
| Grafana panel: the 3D twin as a panel whose time range follows the dashboard | Roadmap |
| Grafana alert rules as DC Sim alarms with advice | Roadmap |
Nothing inbound
The agent will dial out over one WebSocket and never listen on a port: no remote-write, no push gateway, no scrape target to expose. If the link drops, batches spool to disk and replay in order. Exporter credentials stay in the agent’s configuration; DC Sim receives metric values only.
Is the Grafana and Prometheus integration live today?
Not as a live connector. Today a Grafana panel's CSV export goes into the Readings sheet. The Prometheus collector is built into the agent and is not yet tested against live exporters; the Grafana panel is on the roadmap.
What credentials will it need?
Usually none: exporters are typically unauthenticated inside the network. For an endpoint behind a proxy, a bearer token or a read-only username and password, kept in the agent's configuration.
Will it change anything in Prometheus or Grafana?
No. The agent only reads /metrics endpoints; it does not use remote-write, a push gateway or the Grafana API.
My metric names are custom. Will they work?
Known families map automatically; anything else can be passed through with an include pattern and bound to a device by hand.
One Excel file. Your room in 3D.
Download the import template (v1.3, nine sheets, only Devices required) and the filled example from our demo hall. No form, no email needed, no macros in the files.
Fill in the Devices sheet and send it with your walkthrough request; we import it on the call. What each sheet does
