Consumption

A lightweight Linux tool to monitor processes and estimate their power/energy consumption
using CPU utilization, disk I/O, and memory activity as proxies.

---

Table of Contents

• Features
• Installation
• Releases
• Usage Examples
• Algorithm

Features

• Process-level monitoring

  • Accepts single PIDs, multiple PIDs, or ranges (1000..1010).
  • Works with process trees via pstree expansion.

• Multiple output formats

  • Human-readable table (default).
  • CSV and JSON streams for machine processing.
  • Self-contained HTML report with summary and per-tick table.

• Configurable model

  • Idle power, max power, and CPU nonlinearity exponent.
  • Disk energy per byte read/write.
  • Memory RSS churn and refault energy.
  • Adjustable idle-share distribution (--alpha).

• Post-processing tools

  • calc subcommand computes averages from a saved CSV/JSON report.

• Safe defaults

  • Ships with reasonable coefficients for typical laptop/server workloads.
  • All parameters are overrideable via CLI flags.

---

Installation

Install directly from source using go install:

go install github.com/ja7ad/consumption/cmd/consumption@latest

This will place the consumption binary in your $GOPATH/bin or $HOME/go/bin.

This tool requires cgroup v1 or v2 with sudo privileges to read process stats.

---

Releases

Prebuilt binaries are available on the Releases page. Download the latest release for your platform and place it in your $PATH.

---

Usage Examples

Monitor a single process

consumption -s 10 -i 1s $(pidof nginx)

Collects 10 samples at 1 second interval for the nginx process.

---

Monitor a process tree (e.g., JetBrains GoLand IDE)

consumption -s 20 -i 1s $(pstree -p $(pidof goland) | grep -o '([0-9]\+)' | tr -d '()' | tr '\n' ' ')

Expands all child PIDs of GoLand and monitors them together.

---

Continuous monitoring until stopped

consumption -s 0 -i 500ms -- $(pidof postgres)

Runs indefinitely (until Ctrl-C), sampling every 500ms.

---

Save reports to CSV and JSON

consumption --csv out.csv --json out.json -- $(pidof redis-server)

Outputs per-tick rows to both CSV and JSON.

---

Generate an HTML summary report

consumption --html report.html -- $(pidof mysqld)

Produces a full HTML report with averages, energy totals, and per-tick data.

---

Post-process a report file

consumption calc out.csv

Reads an existing CSV/JSON report and calculates average consumption:

consumption avg (over 18 samples of ~1s):
- watt (cpu):    0.013 W
- watt (disk):   0.000 W
- watt (ram):    0.000 W
- watt (total):  0.013 W

---

Algorithm

⚠️ Inside a KVM or cloud VM you cannot read real voltage or exact watts of a single process — the hypervisor owns hardware counters (RAPL/PKG/DRAM).
Instead, Consumption builds a model that estimates process power by combining CPU utilization, I/O, and memory activity with calibrated coefficients.

1. Symbols

• $N$: number of vCPUs
• $\Delta t$: sampling interval (seconds)
• $U_{\text{proc}}$: process CPU utilization fraction
  $U_{\text{proc}}=\dfrac{\text{CPU time of process in sec during }\Delta t}{N \cdot \Delta t}$
• $U_{\text{vm}}$: total VM CPU utilization fraction
  $U_{\text{vm}}=\dfrac{\text{CPU time of all tasks}}{N \cdot \Delta t}$
• $P_{\text{idle}}$: idle power in Watts (configurable)
• $P_{\text{max}}$: max power in Watts at 100% utilization
• $\gamma$: CPU nonlinearity exponent
• Disk bytes: $B_r$ (read), $B_w$ (write)
• Memory proxies:
  • RefaultB = refaulted bytes
  • RSSChurnB = RSS churn bytes
• Energy per byte coefficients:
  • $e_r$ (disk read), $e_w$ (disk write)
  • $e_{\text{ref}}$ (RAM refault), $e_{\text{rss}}$ (RSS churn)

2. VM power model

Dynamic VM power grows nonlinearly with utilization:

$$ P_{\text{dyn}}(U) = (P_{\text{max}} - P_{\text{idle}})\cdot U^\gamma $$

Total VM power at utilization $U_{\text{vm}}$:

$$ P_{\text{vm}}(U_{\text{vm}}) = P_{\text{idle}} + P_{\text{dyn}}(U_{\text{vm}}) $$

3. Attribute CPU power to process

Each process is charged a share of VM dynamic power:

$$ P_{\text{cpu,proc}} = \begin{cases} \dfrac{U_{\text{proc}}}{U_{\text{vm}}} \cdot P_{\text{dyn}}(U_{\text{vm}}), & U_{\text{vm}}>0 \\ 0, & U_{\text{vm}}=0 \end{cases} $$

4. Optional idle power distribution

If policy requires spreading idle cost, an $\alpha$ factor applies:

$$ P_{\text{idle,share}} = \alpha \cdot P_{\text{idle}} \cdot \dfrac{U_{\text{proc}}}{U_{\text{vm}}} $$

• $\alpha=0$: no idle charged (default)
• $\alpha=1$: full idle proportionally shared

5. Disk and memory power

Convert per-byte activity to Joules, then divide by $\Delta t$:

$$ P_{\text{disk}} = \frac{e_r \cdot B_r + e_w \cdot B_w}{\Delta t} $$

$$ P_{\text{ram}} = \frac{e_{\text{ref}} \cdot \text{RefaultB} + e_{\text{rss}} \cdot \text{RSSChurnB}}{\Delta t} $$

6. Total process power and energy

Total instantaneous power:

$$ P_{\text{proc}} = P_{\text{cpu,proc}} + P_{\text{disk}} + P_{\text{ram}} + P_{\text{idle,share}} $$

Cumulative energy (Joules) is the time integral:

$$ E_{\text{proc}} = \sum , P_{\text{proc}} \cdot \Delta t $$

7. Future extensions

• Network I/O: add a term $e_n \cdot N_b$ with $e_n$ = energy per byte transferred.
• Device-specific coefficients: tune $e_r$, $e_w$, $e_{\text{ref}}$, $e_{\text{rss}}$ for different hardware.