open-standards/eoc/eoc-rs
Repository files (latest commit first)
Filename Latest commit message Latest commit date
Transaction Science 26a1b6aa63 eoc-meter: an interval long past the tick is no evidence that a rail is live
The liveness rule counts intervals and never looked at how long one lasted. A
rail that steps twice a second steps in every interval once sampling slows to
that cadence, and a contended host slows it just so. On an M3 Ultra the same
rail, burned and sampled by the same code in one quiet process, was Frozen in 17
of 17 probes ticking at 25, 100 or 250 ms and Live in 12 of 12 at 500 or 800 ms.
With four test binaries sampling at once, 6 of 20 runs cached Live for the
process, on a rail every other crate on that host called Frozen, and in CI the
status selected the counter.

An IOReport interval now counts only within MAX_INTERVAL_TICKS (3) of the tick
the probe asked for. A longer one shows which rails exist, and is otherwise no
evidence either way: not its advances, not the interval, not its time. So the
300 ms window is 300 ms of intervals that count, or a burst beside one long
interval would fill it. A probe that cannot sample near its cadence ends Frozen
and falls back to its estimator: estimated, never over-claimed.

Three, from interval streams recorded on real hosts and replayed under each cap.
That rail under contention: Live in 57 of 80 streams with no cap, 13 at eight
ticks, 9 at five, 2 at four, none at three. A RAPL counter that is live: Live in
10 of 10 on a quiet host, 9 of 10 with a spinning thread for every core, 2 of 10
with twice as many.

Only IOReport's rails have the cap. RAPL, NVML and a process's counter publish
at least as often as their probes tick, so sampling slowed by contention shows
nothing a quiet probe would not. Capping them too made the tests that hold a
process's counter to Live fail under load: 5 of 27 runs on an M5 Max with four
busy threads a core, four of them on the counter's status.

RailWatch::default is IOReport's watch, for a probe ticking at PROBE_TICK.
every_channel, the vendor counters' watch, has no cap. The tests that called a
rail sampled every 600 ms Live, as intended, now say it is not.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-11-07 03:17:48 -05:00
..
crates eoc-meter: an interval long past the tick is no evidence that a rail is live 2026-11-07 03:17:48 -05:00
Cargo.lock sec: answer the gate's first catch — two constant-time advisories, and pin an allowance to the version it excuses 2026-10-24 20:06:05 -04:00
Cargo.toml Cargo: no minimum-supported-version floor, because the policy is latest stable 2026-11-03 17:42:20 -05:00
LICENSE licences: ship the Apache-2.0 text the code already declares 2026-10-09 17:22:32 -04:00
README.md eoc docs: what a resolve is charged, with the measured figures; the example prints the unmetered baseline 2026-10-13 02:22:58 -04:00
rust-toolchain.toml toolchain: Rust 1.98.0 -> 1.99.1 (change-controlled pin) 2026-11-03 17:42:20 -05:00

eoc-rs — EOC reference implementation in Rust

Apache-2.0 reference implementation of the Energy-Optimized Compute specification (CC-BY-4.0 spec lives in ../spec/).

EOC is a four-stage memoizing cascade. Every query is tried at each stage in turn and the first that can answer wins. The four stages, cheapest to most expensive, are:

  1. Cache (eoc-cache) — content-addressed LRU. Cache hits are nearly free.
  2. Key-value (eoc-kv) — exact key match plus cosine-similarity embedding match.
  3. Graph (eoc-graph) — triple-store retrieval. Reference impl is a small in-memory matcher; production deployments swap in a real graph backend driven by DCY (EOC-4).
  4. Neural (eoc-neural) — last-resort inference. Reference impl ships an EchoBackend; production deployments wire in llama.cpp / Ollama / Anthropic API / etc.

Joule cost is the unit of accounting. The eoc-meter crate reads cumulative micro-joules from whatever whole-work hardware counter is available and live (Linux RAPL, Apple Silicon IOReport; NVIDIA NVML behind the cuda feature is GPU telemetry, never the cascade's counter) and falls back to a StubCounter everywhere else — including WASM.

What a resolve is charged

A cost is Measured only when every micro-joule of it was read from a live counter over the work it is booked to; anything else is Estimated. The cascade reads its meter where the meter can resolve the work, and nowhere else.

  • A resolve the neural stage answers is read. Two reads of the meter bracket the stage, the first taken once the cache, the KV stage and the graph have missed. The span is combined with the cost the stage reported: measured when the stage estimated nothing, the larger of the two readings when the stage metered itself, an estimate when the stage estimated. The three misses are spent before the first read and are not in the figure; the meter's own closing cost is. On an M3 Ultra the misses together took 220 ns and 0.95 µJ, and a span over no work read 12 to 29 µJ, so the measured figure is not under the neural stage's energy.
  • A resolve the cache, the KV stage or the graph answers reads no meter, and is never measured. It is charged the stage's own figure plus what a hit on that tier costs on this host. The first time a tier answers, the cascade repeats that lookup inside one span of the meter until the span reads fifty times what a span over no work does (or 40 ms have gone), and keeps the reading per lookup, in nanojoules. Every later hit on the tier is charged that, in whole micro-joules with the remainder carried to the next hit, so the micro-joules charged over any number of hits sum to the calibrated total. Cascade::calibration returns the figure, the batch, the span and the empty span, or why a tier has none; Cascade::calibrate calibrates again; Cascade::without_hit_calibration switches a tier off.

A hit is not bracketed because a read of a counter costs more than a hit. The in-process counter on macOS takes each read only once the reading thread is seen to have left its core, about 80 µs. With a span around every resolve, 10,000 cache hits took 1,647 ms on that host, 164 µs each, and each was sealed as a measured 13 to 28 µJ: the meter's own cost, the same as a span around a neural stage that does nothing. Without it they take 4.2 ms, 0.38 µs each, and are charged the 1.2 to 1.7 µJ a cache hit was calibrated at there (2.6 to 3.4 µJ for a KV or graph hit; the calibration takes 1 to 2.4 ms, once a tier), as an estimate: read by this counter on this host, but over a batch, and the mean of a batch is not a reading of any one hit.

With no live counter there is no calibration and no span, and every resolve keeps the cost its stage reported.

Workspace layout

Crate Purpose
eoc-core Types: Query, Response, Stage, JouleCost, Receipt.
eoc-cache Stage 1 — LRU cache + the Stage trait.
eoc-kv Stage 2 — KV + embedding-similarity lookup.
eoc-graph Stage 3 — in-memory triple store.
eoc-neural Stage 4 — neural inference trait + EchoBackend.
eoc-meter Hardware energy counters with stub fallback.
eoc-cascade Glue: walks the four stages, attributes joule cost.
eoc Reference CLI.
eoc-wasm wasm-bindgen wrapper — browser-runnable cascade.

Running

cargo test --workspace
cargo clippy --workspace --all-targets
cargo run --bin eoc -- query "what is 2+2"
cargo run --bin eoc -- meter

Example output, on a host with no live counter, where the cost is the stage's own figure:

$ cargo run --bin eoc -- query "what is 2+2"
stage  : kv
payload: 4
cost   : 50 µJ (estimated)
hit    : the kv stage's own figure; no calibration by stub (the counter did not advance)
receipt: <64-char hex>

and on a Mac whose in-process counter is live (--release; the stage's 50 µJ plus the calibrated figure):

$ cargo run --release --bin eoc -- query "what is 2+2"
stage  : kv
payload: 4
cost   : 53 µJ (estimated)
hit    : a kv hit is charged 3357 nJ on this host, on top of the stage's own figure: process-energy read 6876 µJ over 2048 lookups in one span (24 µJ over none)
receipt: <64-char hex>

cargo run --release -p eoc-cascade --example hit_cost prints what each tier's hit takes and is charged on the host it runs on.

CUDA / NVML

NVIDIA NVML support is feature-gated and off by default so the workspace builds without libnvidia-ml. NVML's total-energy counter reads the GPU alone, so detect() never hands it to the cascade, whose costs carry no field to say a figure covers part of the work: NvmlCounter is GPU telemetry, read beside the whole-work counter, and it is constructed only while its counter is live and counts no faster than the board can draw. To enable:

cargo build -p eoc-meter --features cuda

WASM

The cascade compiles to wasm32-unknown-unknown. The meter does not — RAPL, NVML, and powermetrics are all absent in the browser, so WASM builds use StubCounter and report estimated joule cost.

rustup target add wasm32-unknown-unknown
cargo build -p eoc-wasm --target wasm32-unknown-unknown --release
wasm-bindgen target/wasm32-unknown-unknown/release/eoc_wasm.wasm \
    --out-dir ./pkg --target web

Then from JavaScript:

import init, { WasmCascade } from "./pkg/eoc_wasm.js";
await init();
const cascade = new WasmCascade();
const r = await cascade.resolve("ping");
console.log(r); // { stage: "kv", payload: "pong", microjoules: 50, ... }

License

Apache-2.0 for the Rust source in this workspace. The EOC specification text in ../spec/ is CC-BY-4.0. See ../LICENSE.