FastAPI middleware: attribute per-request energy by CPU time and exclude idle power

オープン
  • 何か問題がありますか？
#1,428 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
30/100
issue の種類
機能追加
明瞭さ
おおむね明確
活発さ
活発
技術スタック
fastapi, python
領域
backend-api-design, performance

調査の方向性

Start by reading codecarbon/integrations/fastapi/attribution.py, especially _settle, and the middleware introduced by #1203. Trace how async handlers and FastAPI dependant calls are currently measured before assessing the CPU-time and idle-energy design. Done means the stated conservation invariant and CI checks are covered, with the bare-metal validation documented as complete or still unavailable.

索引モデルが issue の本文から書いたものです。

説明

enhancement

Is your feature request related to a problem? Please describe.

The FastAPI middleware from #1203 splits each sampling window's machine energy across the requests in flight, weighted by wall-clock overlap (codecarbon/integrations/fastapi/attribution.py, _settle). That number is an estimated share, and it is biased in four ways:

  1. Idle power is charged to whichever request happens to be in flight. A single 1 ms request in an otherwise idle 15 s window receives the whole window's energy.
  2. A request waiting on I/O is weighted the same as one using the CPU.
  3. In machine mode, energy used by other processes on the host is charged to requests.
  4. GPU energy is not tied to the request that launched the work.

A benchmark on #1203 (uvicorn, oha, all-MiniLM-L6-v2 on CPU) found the middleware's latency cost within noise, so speed is not the concern here. Accuracy is.

Describe the solution you'd like

Charge each request only the energy above idle that its own process used, split by the CPU time the request consumed:

  • Per window and per component: dynamic = max(ΔE − P_idle × width, 0). The remainder goes to an idle_kwh bucket.
  • Idle power: the lower of a rolling minimum of window power and a regression intercept (used only when the fit is good). In CPU load mode it is known analytically (0.1 × TDP).
  • Process share of CPU dynamic energy: Δprocess_time() / Δ busy CPU time from psutil.cpu_times(). The rest goes to other_processes_kwh.
  • Per-request CPU time: async handlers are metered by wrapping the handler coroutine and summing time.thread_time_ns() around each resumption. Sync handlers are metered by wrapping FastAPI's dependant calls that run in the threadpool. CPU time the meter cannot assign goes to process_unattributed_kwh.
  • GPU dynamic energy stays split by wall-clock overlap and is reported separately, labelled as such.
  • An exact conservation invariant: attributed + idle + other_processes + process_unattributed + unattributed == settled.
  • New RequestEnergy fields: cpu_seconds, attribution_method (cpu_time / wall / mixed), quality (measured for RAPL/powermetrics/NVML, modeled for load mode, none). energy_kwh becomes the dynamic energy. It has not been released yet, so this breaks no users.

Validation, against known ground truth:

Check Threshold Where
Conservation with fake clocks and energy relative error < 1e-12 CI
cpu_seconds of a known CPU burn ±10% CI
Energy ratios of 1:2:4 ms CPU endpoints, concurrency 16 ±10% Linux with RAPL
asyncio.sleep endpoint energy ≤ 2% of the 4 ms burn Linux with RAPL
Per-request energy with a CPU hog in another process changes ≤ 15% Linux with RAPL

Per-request overhead budget: ≤ 30 µs p50 on Linux, ≤ 50 µs on macOS.

Describe alternatives you've considered

  • Idle subtraction only, keeping wall-clock weighting. Simpler, but a request waiting on a database is still charged like one using the CPU.
  • Per-route aggregates only, no per-request numbers. Avoids misreading single values, but does not improve the estimate.
  • Metering child tasks (create_task, gather) with a custom task factory. Left out for now: it can clash with a user's own factory, and the process_unattributed_kwh bucket shows how much CPU time goes unmetered without it.

Additional context

Known limits: CPU time is a proxy for energy, and frequency scaling, SMT and AVX add error (an estimated ±15%, not yet measured). Per-request GPU energy is not possible without instrumenting user code. Wrapping dependant calls relies on FastAPI internals.

The bare-metal checks (the last three rows of the table) have not been run yet: they need a Linux machine with readable RAPL, which GitHub runners likely do not expose.

Related: #1203 (the middleware), #1178 (earlier per-request inference proposal), #936 (several processes sharing one machine).

主要言語
Python
スター
1.9k
フォーク
351
平均マージ
3日 12時間
マージ済み PR（30日）
22

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

mlco2/codecarbon のほかの issue

mlco2/codecarbon の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。