Skip to content

Automatic energy

Fabrizio Degni edited this page Nov 15, 2026 · 3 revisions

Automatic energy collection

NodeLift samples supported power sources automatically while its node or host process is running. No extra pairing step, administrator elevation, model download, Python installation or account is required for the bundled Windows/Android companions. The existing authenticated hub association carries the readings. A collector does not grant permission to execute work.

Supported sources and electrical boundaries

Platform/source Classification What is covered Limit
macOS AppleSmartBattery IORegistry Measured, derived from reported current and voltage Gross battery-terminal discharge of the laptop Only without external power and while discharging. Not wall input or application-only energy. Registry availability varies by OS/device. Unknown calibration.
Android BatteryManager + battery voltage Measured, derived from reported current and voltage Gross battery-terminal discharge of the phone/tablet Negative current, unplugged and a compatible battery status required. Full/not-charging status does not imply an external supply. Unsupported OEM values stay unavailable. Not usable as whole-device consumption while charging.
NVIDIA driver nvidia-smi power.draw, Windows/Linux Measured, driver-reported GPU board only Driver/tool and sensor support required. The driver may return instantaneous or recent averaged power. No inference qualification, calibration or CPU/whole-PC coverage follows.
Linux powercap package RAPL energy counter Estimated CPU package only Read permission required. No elevation is attempted. Model-based counter values remain estimates, including when exposed as hardware counters.
iOS public UIKit APIs Unavailable for W/J Battery charge percentage/state only No fabricated watts from charge percentage, nominal capacity or CPU use. Instruments/MetricKit diagnostics are separate from a live whole-device power meter.
Xbox / unsupported desktop, GPU or ARM devices Unavailable until a supported source exists No assumed coverage Inventory, TDP, theoretical FLOP/s and battery percentage are not energy measurements.

The operating-system/driver reading is the source, not an independent calibrated meter. In particular, “measured” identifies the sensor-backed acquisition path, not a guarantee of laboratory accuracy. Sensor estimation and calibration limitations remain visible. NodeLift does not install a vendor tool, request a new privileged helper, or use private iOS APIs automatically.

Hive presentation

The Overview shows measured and estimated power subtotals separately, per-node source/scope, sample time, uncertainty, recent power traces, common-window Wh and coverage. Every enrolled resource remains represented when offline, unsupported or stale. A measured node's percentage is its share of the measured subtotal, not its share of an unobserved entire estate.

These are gross node/component readings. Other applications, operating-system activity and the collector itself are included. There is no automatic idle-baseline subtraction, task attribution, energy-saving percentage, billing estimate, carbon estimate, or joules/token claim. Network, cooling and the hub's own hardware are excluded unless separately instrumented. A worker host may appear as an ordinary instrumented node; the Hive service is not silently treated as a metered device.

Aggregation rules

  • Default sampling is once every five seconds. A maximum of 61 points per channel is retained locally, with at most 300 seconds in an outbound report. Raw readings are bounded; local serial numbers, filesystem paths and credentials are not included.
  • Power is a sum of latest fresh readings, each no older than 45 seconds. This is not an exactly simultaneous power measurement. Offline/stale/unknown resources do not become zero-watt nodes.
  • Energy uses at most 60 seconds of a common covered window within each classification. It interpolates bounded sensor samples without extrapolation. A gap over 15 seconds or any missing bracket makes that subtotal's energy unavailable, rather than dropping the problematic channel silently. The window may move or shrink; it is not lifetime consumption.
  • RAPL interval means use right-interval integration rather than pretending to be instantaneous readings. Counter rollback/wrap handling is bounded by the collection interval. Sensors are not polled at benchmark-kernel frequency.
  • A whole-device input/discharge channel shadows component channels on the same resource. Separate CPU-package and discrete GPU-board domains may form an explicitly partial component subtotal. A battery or wall meter is never added to those same device components.
  • Measured and estimated values have separate coverage, windows and totals. They are never silently combined. A missing sensor is not filled by TDP or a nominal battery assumption.
  • Time coordinates are reported device wall clocks with monotonic collection checks. Local clock discontinuities rebuild the window. Cross-device clock accuracy and network delay are not independently attested; the UI states this limit. Freshness checks reject future/old malformed evidence, but do not prove synchronized clocks.
  • Only server-owned resource aliases deduplicate multiple enrolled views. Authentication is attribution of a report, not global physical-device or sensor attestation.

nodelift.energy-telemetry.v1 is optional in authenticated capabilities, dedicated telemetry requests and compute poll/progress requests. Dedicated telemetry never leases work and remains available without CPU admission. The hub binds the node identity from the authenticated session, stores the newest compute-channel report and ignores older report replays. Telemetry persists independently of CPU result qualification and continues on supported workers during benchmark lease renewals. Malformed optional sensor data cannot qualify or invalidate CPU work.

Inventory-forwarded LAN nodes and directly enrolled remote nodes use the same sensor contract. Readings are attributed to their actual enrolled hub/group membership. There is no inference that a node is local based on its IP, nor that four planned regions are operational. Cross-region federation and remote-network benchmark emulation remain separate work.

Onboarding and operation

  1. Install the architecture-appropriate companion, then use Add a device in Hive and scan/open its invitation.
  2. Resume the saved association. Sensor probing starts with the node session; there is no separate sensor setup wizard for supported unprivileged sources.
  3. Open Hive's Energy and consumption area. A few samples are needed before Wh is available. Check the source and electrical scope before interpreting the subtotal.
  4. Android/macOS battery readings require discharge. Connecting external power deliberately makes this boundary unavailable, rather than reporting the charging current as consumption. Windows NVIDIA collection requires an already functioning supported driver and nvidia-smi.
  5. End the Android node session or exit the desktop node process to stop its collector. Stopping only a CPU job keeps desktop presence and energy monitoring active. The local coordinator's collector follows that host process. No arbitrary network sensor scanning or credential provisioning is added.

Validation and remaining work

Automated tests cover units/sign, charging exclusion, measured/estimated separation, alias/domain overlap, common windows, gaps, stale/offline/future/invalid reports, RAPL wrap, clock shifts and authenticated direct/coordinator routing. A physical Mac battery sensor run is recorded separately in the evidence folder. It is not an external wall-meter calibration or a test of other devices.

Android and Windows compilation/package validation do not establish physical sensor availability on a specific phone, GPU or driver. iOS explicitly reports the public-API limit. External metering, calibrated error bounds, independently bounded cross-node clock alignment, long-term time-series storage and model/job-specific incremental energy require additional implementation and validation.

Primary references

Android sensor diagnostics in 0.2.27

The owner-only telemetry and local diagnostic export retain allowlisted battery state, plugged flags, voltage (mV), instantaneous current and average current (µA). Missing properties are explicit. Raw values are diagnostic context, not automatic energy estimates. An instantaneous current sign conflict is never hidden with absolute-value conversion or an average-current substitution. Distinct reasons identify missing current, zero/resolution, invalid voltage, uncertain state, external power and read failures.

An actual OPPO PHY110 session on 15 November 2026 reported USB power and FULL state while connected to the Mac. Removing the wall charger alone did not remove that supply. After the USB cable was physically disconnected, 0.2.26 changed to a generic unavailable sensor reason. A subsequent physical 0.2.27/build 31 session delivered five fresh reports with no external power, discharging status, 4348 mV and strictly positive instantaneous current (+61 to +83 in the API-documented microampere field); average current was unavailable. Hive displayed the sign-conflict reason and raw diagnostics. This establishes the immediate API-level exclusion cause, not a verified vendor polarity/unit correction. OPPO W/J remain unavailable until an independent counter or meter can validate a usable acquisition path.

Clone this wiki locally