logos-blockchain/logos-blockchain-pocs
1
0
Fork 0
mirror of https://github.com/logos-blockchain/logos-blockchain-pocs.git synced 2026-07-04 22:39:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

pq-bench-rpi5: consolidated PQ benchmark snapshot (KEM + sig + TLS)

Browse Source 
Post-quantum primitive benchmark harness with two consolidated baselines,
each measuring KEM + signatures + TLS 1.3 handshakes in one pass:

- RPi5 (cortex-a76, baseline-grade): the validator reference baseline
  (performance governor, core-pinned, no throttling).
- Apple M3 (apple-m3, is_baseline_grade=false by design): cross-platform
  reference, full reps=5 auto-calibrated run.

Candidates: ML-KEM 512/768/1024, Classic McEliece (6 sets) and FrodoKEM
(3 AES variants), ML-DSA, Falcon, SLH-DSA, plus the X25519/Ed25519 classical
baseline and a TLS hybrid-group x PQ-signature handshake matrix. Both runs
share liboqs 0.15.0 (97f6b86) + oqs-provider 0.9.0; provenance is stamped
into each results JSON. Includes the static dashboard with the merged dataset.

Supersedes the earlier add-pq-bench-rpi5 import.
This commit is contained in:
megonen 2026-06-26 23:07:33 +03:00
parent e0a1176789
commit 7fa0f6fbcb
27 changed files with 32718 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
pq-bench-rpi5/.gitignore vendored Normal file
View File

@ -0,0 +1,41 @@
# Build trees / vendored deps (rebuilt by setup/setup.sh)
/vendor/
/build/
setup/versions.lock
# Compiled harness binaries
bench/kem_sig/bench_pq
bench/tls/bench_tls
*.o
# Generated TLS material
bench/tls/pki/
# Per-run scratch dirs (intermediate artifacts; never committed)
results/.work-*/
# Local results are kept by users, not committed by default.
# Comment the next line out if you DO want to commit your machine's results.
results/*.json
# Consolidated baseline-grade RPi5 run (KEM + sig + TLS in one pass): the Pi
# validator reference baseline.
!results/rasberrypi5-20260625T202356Z.json
# Consolidated Apple M3 cross-platform run (apple-m3, KEM + sig + TLS in one
# pass): is_baseline_grade=false by design, kept as the cross-platform reference.
!results/mehmetmac-20260625T220618Z.json
# Merged dataset consumed by the dashboard. Tracked: ships the published
# baselines so a fresh clone renders charts out-of-the-box. Contributors who
# re-merge overwrite it locally (modified file they can commit or discard).
analyze/png/
# Python venv (matplotlib lives here; system python stays clean)
analyze/.venv/
# Machine-local Claude Code settings (never shared)
.claude/settings.local.json
# Python / OS cruft
__pycache__/
*.pyc
.DS_Store

41
pq-bench-rpi5/Dockerfile Normal file
View File

@ -0,0 +1,41 @@
# =============================================================================
# Reproducible BUILD of the PQ benchmark toolchain on Debian aarch64 (the same
# OS family as Raspberry Pi OS / Ubuntu on the RPi5).
#
# This image is for BUILDING ONLY — it pins and compiles liboqs / OpenSSL /
# oqs-provider reproducibly. It is NOT for running the benchmark.
#
# docker build -t pq-bench-rpi5 . # build + pin the toolchain
#
# Run the MEASUREMENT bare-metal on the host, never in the container. A
# container cannot reliably control the CPU governor, pin to an isolated core,
# or read the Pi's SoC thermal/throttle sensors (vcgencmd) — the three knobs the
# reference-grade gate depends on — so an in-container run could never be
# baseline-grade and would only add noise. Build here if you like; then:
#
# ./run.sh # on the host (see README)
# =============================================================================
FROM debian:bookworm-slim
ENV DEBIAN_FRONTEND=noninteractive
RUN apt-get update && apt-get install -y --no-install-recommends \
build-essential cmake ninja-build git perl pkg-config \
python3 python3-venv ca-certificates \
libssl-dev cpufrequtils util-linux \
&& rm -rf /var/lib/apt/lists/*
WORKDIR /app
COPY . /app
# Build + pin the toolchain at image-build time so the image is self-contained.
# (Comment this out to keep the image thin and run setup.sh at container start.)
RUN ./setup/setup.sh all || (echo "setup failed — see log above" && exit 1)
# Optional: matplotlib PNG export in an isolated venv.
RUN python3 -m venv analyze/.venv \
&& analyze/.venv/bin/pip install --no-cache-dir -r analyze/requirements.txt
ENTRYPOINT ["/bin/bash", "-lc"]
# This image builds the toolchain; it does not run the benchmark. The default
# command just says so — run the measurement bare-metal on the host (see README).
CMD ["echo 'Toolchain built. Run the benchmark BARE-METAL on the host (./run.sh) — Docker is for reproducible builds only; a container cannot meet the baseline-grade gate.'"]

424
pq-bench-rpi5/README.md Normal file
View File

@ -0,0 +1,424 @@
# pq-bench-rpi5
A reproducible, general-purpose **post-quantum cryptography benchmark** whose
baseline target is the **Raspberry Pi 5** (Broadcom BCM2712, Cortex-A76,
aarch64). Anyone can run it on their own Pi 5 and the results aggregate and
compare apples-to-apples.
**Framing — migration cost.** How much does moving from the cryptography Logos
uses *today* (X25519 key exchange + Ed25519 signatures) to PQ candidates cost on
validator-grade hardware? Every chart draws that classical baseline as the
reference line, so the PQ "tax" is always visible.
Phase 1 covers **PQ KEMs**, **PQ signatures**, and **PQ TLS 1.3 handshakes**.
Hooks are left for a later SNARK/STARK phase (see `config.yaml`); it is not
implemented yet.
---
## What gets measured
| Layer | Metrics |
|-------|---------|
| **KEM** | keygen / encaps / decaps wall-clock (median, MAD, IQR, min, max, mean, stddev, ops/sec) · pk/sk/ct sizes · heap high-water |
| **Signature** | keygen / sign / verify wall-clock (same stats) · pk/sig sizes |
| **TLS 1.3** | full-handshake latency · handshakes/sec · bytes-on-wire · ClientHello size (+ fragmentation flag) — as a matrix of (KEM group × signature) |
The **classical baseline** (X25519 / Ed25519 / X25519+Ed25519) is always
included as the reference point — measured as a real primitive via OpenSSL, not
hand-waved.
---
## Project layout
```
pq-bench-rpi5/
setup/ build + pin liboqs, OpenSSL 3.5+, oqs-provider (versions.env / versions.lock)
bench/kem_sig/ bench_pq.c primitive KEM/sig harness (liboqs + OpenSSL EVP baselines)
bench/tls/ bench_tls.c in-process TLS 1.3 handshake harness (OpenSSL API)
run_tls.sh PKI generation + (KEM × sig) matrix driver
bench/lib/ assemble.py / merge helpers / miniyaml.py (zero-dep YAML)
results/ results/<host>-<timestamp>.json (one per run, full metadata)
analyze/ merge.py (combine machines) + plot.py (matplotlib PNGs, optional venv)
dashboard/ static HTML/JS (Chart.js) — no backend, GitHub-Pages deployable
run.sh governor + taskset + thermal wrapper + orchestrator
config.yaml candidate lists (extend here)
Dockerfile reproducible Debian-aarch64 build
```
---
## Quick start
### On a Raspberry Pi 5 (the real measurement target)
```bash
git clone <this repo> && cd pq-bench-rpi5
./setup/setup.sh # build + pin liboqs, OpenSSL 3.5+, oqs-provider
sudo ./run.sh # sudo only to set the performance governor (see below)
python3 analyze/merge.py results/*.json -o dashboard/data/merged.json
# open dashboard/index.html (or deploy dashboard/ to GitHub Pages)
```
**On `sudo`:** it is **optional, not a prerequisite.** The only thing it does is
set the CPU governor to `performance` — none of the crypto needs root. `./run.sh`
runs fine without it: it warns, skips the governor step, completes the run, and
the results JSON is automatically stamped `is_baseline_grade=false` (governor
demerit). So use `sudo` when you want a baseline-grade reference run; drop it for
a quick local run you don't intend to submit.
`./run.sh --smoke` runs tiny iteration counts as a fast pipeline check.
`./run.sh --kemsig-only` / `--tls-only` scope the run. `--iters/--warmup/--reps`
override the `config.yaml` knobs.
### On macOS (development / smoke testing only)
```bash
brew install cmake openssl@3 git
./setup/setup.sh
./run.sh --smoke # produces valid JSON; stamped is_baseline_grade=false
```
> **macOS runs are cross-platform / smoke data, never baseline-grade — by
> design, for three concrete reasons:**
> 1. **Not a Raspberry Pi**, so it fails the gate's first condition outright.
> 2. **No userspace cycle counter, and ~1 µs timer granularity.** macOS exposes
> no readable PMU cycle counter and its wall-clock quantizes to ~1 µs steps —
> a ~10% floor on the fastest ops (ML-KEM ~10 µs), negligible for anything
> ≥100 µs (McEliece, FrodoKEM). (See "Timing source" above.)
> 3. **No Linux cpufreq governor, and core-pinning isn't guaranteed.** Two of the
> noise-control knobs the gate relies on — `performance` governor and a pinned
> core — aren't available, and the build flags aren't `cortex-a76` either.
>
> Every macOS results file records `is_baseline_grade=false` with the exact
> reasons, and the dashboard hides such runs by default. They still produce
> **useful cross-platform numbers** (the heavier McEliece/FrodoKEM ops are barely
> affected by the timer floor) — they just can't meet the controlled reference
> bar, hence smoke-only.
### Docker (reproducible build — build only, never run)
Docker is for reproducibly **building** the pinned toolchain (liboqs / OpenSSL /
oqs-provider), not for running the benchmark:
```bash
docker build -t pq-bench-rpi5 . # builds + pins the toolchain inside the image
```
**Run the measurement bare-metal on the host.** A container can't reliably set
the CPU governor, pin to an isolated core, or read the Pi's thermal/throttle
sensors — the noise-control knobs the reference-grade gate relies on — so an
in-container run could never be baseline-grade and would only add jitter. Build
in Docker if you like; then run `./run.sh` on the host.
---
## Measurement methodology (why the numbers are credible)
`run.sh` is the wrapper that makes a number defensible:
- **CPU governor → `performance`** (Linux; needs `sudo`). Recorded before/after.
If it can't be set (e.g. not root) the run **continues anyway**: it warns,
proceeds, and the missing governor becomes an `is_baseline_grade=false`
demerit. `sudo` is only ever for this step — never for the crypto.
- **Core pinning via `taskset -c 3`.** This is a **single-operation latency**
benchmark (one keygen, one encaps, one sign — timed in isolation), not a
parallel-throughput one, so pinning the whole sweep to one core keeps that
core's cache warm and removes cross-core migration scheduling noise, which
tightens the median and MAD. The Pi 5 has 4 cores (0–3); core **3** is chosen
because core 0 typically absorbs the most OS/IRQ/RPS work. The pinned core and
exact `taskset` command are recorded.
- *Planned (separate axis):* a multi-core **throughput/scaling** mode — run an
op across 1..N cores and report ops/sec plus scaling efficiency per
algorithm. Some schemes (SLH-DSA, and later STARK proving) parallelize far
better than others, so it's a worthwhile dimension — but kept **separate**
from these per-op latency numbers, not mixed into them.
- **Thermal/clock trace.** A background sampler logs ARM clock
(`vcgencmd measure_clock arm`) and SoC temperature (`vcgencmd measure_temp`)
~once a second for the whole run. The full trace is embedded in the results
JSON, and **thermal throttling** (`vcgencmd get_throttled`, plus a clock-droop
heuristic) is detected and flagged — a throttled run is not baseline-grade.
- **Warmup + N timed iterations, multiple repetitions.** Primary metric is
wall-clock nanoseconds via `clock_gettime(CLOCK_MONOTONIC)`. We report
**median, MAD, IQR, min, max, mean, stddev, ops/sec**, plus per-repetition
medians — not just a mean.
- **Timing source — two clocks, honestly recorded.** There are two ways to time
an op:
1. **Cycle-based** via the ARM hardware cycle counter (`PMCCNTR_EL0`) — the
most precise, but on Linux **userspace can't read it by default**: the
register traps unless a kernel module enables the userspace PMU (e.g.
`enable_arm_pmu`).
2. **Time-based** wall-clock via `clock_gettime(CLOCK_MONOTONIC)` — always
available, and accurate enough for the millisecond/microsecond ranges here.
The harness probes the cycle counter and, when it isn't available, **falls
back to wall-clock and records exactly that** in the JSON
(`run.cycles_available=false` + the reason). **On a stock machine the cycle
counter is not available, so runs use the wall-clock timer by default** — and
both published runs reflect this: the RPi5 baseline and the macOS run *both*
have `cycles_available=false` (both wall-clock). The remaining difference
between them is wall-clock **granularity**, not clock *type*: the Pi's
wall-clock lands on fractional microseconds, while macOS quantizes to ~1 µs
steps — a ~10% resolution floor on the fastest ops (ML-KEM keygen ~10 µs),
negligible for anything ≥100 µs (McEliece, FrodoKEM).
- **CPU features / Keccak acceleration.** NEON, SHA2, SHA3, SHA512, AES, PMULL
are detected (`/proc/cpuinfo` on Linux, `sysctl` on macOS). **Note:** the
Cortex-A76 has the SHA2/AES extensions but **not** the ARMv8.2 SHA3
extension, so on the Pi 5 Keccak runs on NEON/scalar code — the results record
both the hardware capability and whether liboqs was compiled with SHA3
instructions, so this is explicit rather than assumed.
### The AArch64-optimized backend
liboqs is built with `OQS_DIST_BUILD=OFF` and the pinned flags so the optimized
aarch64 ML-KEM backend (`mlkem-native`) and Falcon/Keccak asm are compiled in.
`setup/setup.sh` extracts the proof from the generated `oqsconfig.h` (e.g.
`OQS_ENABLE_KEM_ml_kem_768_aarch64 1`) into `versions.lock`, which is stamped
into every results file under `toolchain.liboqs_opt_defines`.
---
## Methodology & trustworthiness (verify it yourself)
Every claim below points at the exact code so you can read it, not take our word.
All `bench_pq.c` references are `bench/kem_sig/bench_pq.c`.
1. **Correctness gate — broken crypto emits *zero* numbers.** Before any timing,
each algorithm runs a full round-trip and asserts it: for KEM,
keygen→encaps→decaps then `memcmp(ss_encaps, ss_decaps)`
(`bench_pq.c:357-363`); for signatures, keygen→sign→`verify` must succeed
(`bench_pq.c:428-434`). On any failure, `die()` prints to **stderr** and
`exit(3)` (`bench_pq.c:303-307`) — and the JSON is only printed *after* all
measurement (`bench_pq.c:372-381`), so a failed gate yields **no stdout at
all**. The gate runs once, *outside* the timed loop. A runtime guard
(`must_measure`, `bench_pq.c:311-315`) also aborts if a timed op ever fails
mid-run. *Verify it:* flip one byte of the decaps shared secret right before
`bench_pq.c:362`, rebuild, run — the process exits `3` with empty stdout.
2. **No dead-code elimination — the `volatile` sink.** At `-O3` the compiler may
delete work whose result is never observed. Each timed op folds an output
byte into a file-scope `volatile uint64_t g_sink` (`bench_pq.c:300`; uses at
`:333,:336,:339,:407,:410,:486`), forcing the store to be materialized so the
crypto call **cannot** be optimized away. Without it the loop could time
nothing and report meaningless near-zero numbers.
3. **What is timed — only the op, never setup.** The timed region brackets a
single `fn(ctx)` call between two `now_ns()` reads (`bench_pq.c:274-281`);
per-rep warmup runs *outside* it (`bench_pq.c:272-273`). Inputs are canonical
and pre-validated, so e.g. KEM decaps (`bench_pq.c:337-339`) times one
`OQS_KEM_decaps` and nothing else. For the X25519 baseline, keygen is timed
separately (`bench_pq.c:507`), a stable key is re-primed *outside* timing
(`bench_pq.c:509`), then derive is timed alone (`bench_pq.c:510`) — setup is
never folded into a measured number.
4. **Per-op auto-calibration with clamps.** `calibrate_op` (`bench_pq.c:209-250`)
runs a doubling probe (`:223-230`, also cache warmup) to estimate per-op cost
`est_ns` (`:231`), then picks iterations to hit `target_time_ms` of real work
(`:234-235`), clamped to `[min_samples, max_iters]` (`:236-237`). So a fast
18 µs keygen and a 0.74 s SLH-DSA sign each get the iteration count *they*
need: slow ops floor at `min_samples` (30), fast ops ceil at `max_iters`
(20000). The chosen `timed_iters` and `calib_est_ns` are recorded per op.
5. **Robust statistics — median + MAD.** `compute_stats` (`bench_pq.c:111-146`)
reports median, MAD, IQR, q1/q3, min, max, mean, stddev, ops/sec, plus
per-repetition medians (`print_stats_json`, `bench_pq.c:184-203`). The
headline metric is the **median**, with **MAD** as spread: timing
distributions are right-skewed with a hard floor (true cost) and a long tail
of OS-scheduling/interrupt contamination that drags mean/stddev but not
median/MAD. Mean and stddev are kept in the JSON so the skew is visible. The
clock is `clock_gettime(CLOCK_MONOTONIC)` (`bench_pq.c:44-48`); userspace PMU
cycles are probed and honestly reported absent when they trap
(`probe_pmu`, `bench_pq.c:66-86`).
6. **`is_baseline_grade` demerit gate.** Computed in
`bench/lib/assemble.py:155-168` as a demerit accumulator — the flag is `true`
only if *every* condition holds: real Pi (`:157`), `performance` governor
(`:160`), core-pinned (`:162`), `cortex-a76` build flags (`:164`), and no
thermal throttling (`:166`). Throttling is read from `vcgencmd get_throttled`
bits 2/18 plus a clock-droop heuristic (`assemble.py:91-98,:110-113`). Any
failure appends a human-readable reason and flips the flag to `false`; the
dashboard and `plot.py` default to baseline-grade runs only.
---
## Reproducibility & provenance
- **Pinned versions** live in `setup/versions.env` (liboqs `0.15.0`, OpenSSL
`3.5.0`/`≥3.5`, oqs-provider `0.9.0`). After cloning, `setup.sh` records the
**actually resolved git commits** and the **exact build flags + compiler
version** into `setup/versions.lock`.
- **Every results JSON carries full environment metadata**: RPi model, RAM,
kernel, OS, governor, the clock/temp trace during the run, compiler version,
liboqs/oqs-provider/OpenSSL versions+commits, build flags, and the candidate
list. A macOS smoke file and an RPi5 baseline file can never be confused.
- **Identical flags for every candidate:** `-O3 -mcpu=cortex-a76` on the Pi.
Document your `gcc`/`clang` version — it is auto-captured in `versions.lock`
(`CC_VERSION`).
### `is_baseline_grade`
A **reference-measurement quality gate**, not a deployment requirement. It marks
whether a run was produced under controlled, reproducible *reference* conditions,
so the numbers are comparable across algorithms and across machines. It is `true`
**only** when all hold: real Raspberry Pi · `performance` governor · core-pinned ·
`cortex-a76` build flags · no thermal throttling. Otherwise it is `false` with a
list of reasons.
- **What it is:** a label that says "this run is clean enough to sit in the
cross-algorithm / cross-machine reference comparison." The dashboard and
`plot.py` default to baseline-grade runs only, so noisy runs don't distort the
picture.
- **What it is *not*:** a claim about how nodes must be configured in production.
Real deployments are heterogeneous (different SoCs, governors, thermals) —
that's a separate question this flag does not speak to.
- A run that doesn't meet the gate **isn't wrong** — it's just flagged
`is_baseline_grade=false` with the reasons and kept out of the reference set.
The macOS cross-platform runs are exactly this: useful, honest numbers that
simply aren't reference-grade.
---
## Candidates (edit `config.yaml`)
- **KEM:** ML-KEM-512/768/1024; hybrids X25519MLKEM768, SecP256r1MLKEM768
(hybrids are benchmarked in the TLS layer; at the primitive layer liboqs
exposes them only as TLS groups, so they show as `enabled:false` there).
Code-based + conservative-LWE backups: Classic McEliece
348864/460896/460896f/6688128/6960119/8192128 (tiny ciphertext, slow keygen)
and FrodoKEM 640/976/1344-AES (unstructured LWE). Baseline: **X25519**.
- **Signatures:** ML-DSA-44/65/87; SLH-DSA (SPHINCS+) variants;
Falcon/FN-DSA-512/1024. Baseline: **Ed25519**.
- **TLS:** matrix of configured KEM groups × signature algorithms, always
including the classical **X25519 + Ed25519** pair.
Classic McEliece and FrodoKEM are now measured (above). **HQC** is not — it is
not enabled in the linked liboqs 0.15.0 build (disabled upstream after the
IND-CCA2 implementation issue), so it is intentionally omitted rather than
listed-and-disabled; re-add it once linked against a liboqs that re-enables it.
Add further algorithms by uncommenting/adding entries — the harness skips
anything your liboqs build doesn't enable (and says so).
---
## Output & analysis
- `results/<hostname>-<timestamp>.json` — one self-describing file per run.
- `analyze/merge.py results/*.json -o dashboard/data/merged.json` — merge runs
from many machines into one dataset (keeps each run distinct; never mixes
baseline with smoke).
- `analyze/plot.py` — matplotlib PNGs for papers (optional; install into
`analyze/.venv` via `analyze/requirements.txt` to keep system python clean —
it gracefully skips if matplotlib is absent).
- `dashboard/` — static, no-backend dashboard: grouped bars by security level,
size-vs-speed scatter, TLS handshakes/sec, and ClientHello size — each with
the classical baseline drawn as a reference line. Deploy the folder to GitHub
Pages, or open `index.html` via any static server.
---
## Contributing your RPi5 results
The whole point is a **shared, aggregated baseline**: the more Raspberry Pi 5
results we collect under identical conditions, the more confident the migration-
cost picture. If you have a Pi 5, please contribute a run — it takes one command
and a pull request.
### 1. Run under baseline conditions
For your numbers to count as baseline-grade, the run must satisfy the
`is_baseline_grade` gate (real Pi 5 · `performance` governor · core-pinned ·
`cortex-a76` flags · no thermal throttling). To give it the best shot:
- **Use a Raspberry Pi 5** with active cooling (the official Active Cooler or a
fan). PQ signing (esp. SLH-DSA) runs the core hot for a while; without cooling
you *will* throttle and the run is flagged non-baseline.
- **Use the official 27 W USB-C PSU.** Under-voltage also trips the throttle flag.
- **Run on a quiet machine** (close other workloads) so core 3 stays clean.
- **Don't edit `config.yaml`'s candidate list** if you want your run to be
directly comparable to others. (Extending it is fine — just say so in your PR;
extra algorithms simply add columns.)
```bash
git clone <this repo> && cd pq-bench-rpi5
./setup/setup.sh # build + pin liboqs / OpenSSL 3.5+ / oqs-provider
sudo ./run.sh # sudo lets it set the performance governor
```
A full run takes a while (SLH-DSA signing dominates). To check the pipeline
first without committing to the full run, use `sudo ./run.sh --smoke` — but only
a **full** run (not `--smoke`) counts as a submission.
### 2. Confirm it's baseline-grade
When the run finishes, the summary prints `baseline-grade (RPi5): True`. Verify
in the JSON too:
```bash
f=$(ls -t results/*.json | head -1)
python3 -c "import json;d=json.load(open('$f'));print('baseline_grade:',d['is_baseline_grade']);\
print('reasons:',d['baseline_grade_reasons']);\
print('throttled:',d['thermal_trace']['throttling_detected']);\
print('aarch64 ML-KEM backend:', 'ml_kem_768_aarch64 1' in d['toolchain']['liboqs_opt_defines'])"
```
You want `baseline_grade: True`, `reasons: []`, `throttled: False`, and the
backend line `True`. If `is_baseline_grade` is false, the printed reasons tell
you what to fix (usually cooling/PSU/governor) — fix and re-run.
### 3. Submit it
Your `results/<hostname>-<timestamp>.json` is fully self-describing (host model,
kernel, OS, governor, clock/temp trace, compiler + liboqs/oqs-provider/OpenSSL
commits, build flags). It contains your **hostname** and Pi model and nothing
else identifying — if you'd rather not share the hostname, set a name first with
`HOSTNAME=mypi5 sudo ./run.sh`, or just rename the file before submitting.
`results/*.json` is git-ignored by default (so you never accidentally commit
local experiments), so add yours explicitly:
```bash
git checkout -b results/<your-handle>-pi5
git add -f results/<hostname>-<timestamp>.json
git commit -m "results: RPi5 baseline from <your-handle>"
# push to your fork and open a PR
```
**PR checklist** (maintainers will look for these):
- [ ] `is_baseline_grade: true` with empty `baseline_grade_reasons`
- [ ] `thermal_trace.throttling_detected: false`
- [ ] `host.is_rpi: true` and `host.rpi_model` mentions "Raspberry Pi 5"
- [ ] `run.governor_after: performance` and `run.pinned: true`
- [ ] `toolchain.cflags_target: cortex-a76`
- [ ] full run (not `--smoke`): `run.timed_iters` is the `config.yaml` value, not 25
- [ ] unmodified candidate list (or extensions noted in the PR description)
Once merged, your file joins `results/`; anyone can regenerate the aggregated
dataset and dashboard with
`python3 analyze/merge.py results/*.json -o dashboard/data/merged.json`. The
dashboard's run selector will then include your Pi alongside everyone else's.
> Prefer not to use GitHub? Open an issue and attach the JSON file instead — a
> maintainer will add it.
---
## Limitations
- **macOS is smoke-only** (see above): coarse timer, no governor/pinning,
fallback flags.
- **Userspace cycle counts** require a kernel PMU module; default is time-based.
- **Heap/stack memory** is best-effort (`mallinfo2` on glibc; reported
unavailable elsewhere); pk/sk/ct/sig **sizes** are authoritative.
- **TLS handshakes are in-process over memory BIOs** — this isolates crypto
cost cleanly (no socket/scheduler noise) but is not a network RTT model;
ClientHello fragmentation is flagged against a typical 1400-byte MSS.
- **Docker is build-only.** The benchmark is not run in a container — a
container can't reliably control the governor, core pinning, or throttle
detection, so measurement runs bare-metal on the host (see the Docker section).
## Future phase (not implemented)
`config.yaml` reserves a `zk:` section for SNARK/STARK proving/verification
benchmarks; the results schema and dashboard are structured to absorb it later.

86
pq-bench-rpi5/RUNNING-ON-YOUR-RPI5.md Normal file
View File

@ -0,0 +1,86 @@
# Running pq-bench-rpi5 on Your Own Raspberry Pi 5
This benchmark measures post-quantum KEMs, signatures, and TLS 1.3 handshakes
against the classical baseline Logos uses today (X25519 / Ed25519), so every
chart shows the **migration cost** of moving to PQ on validator-grade hardware.
There's no manual tuning: the benchmark **auto-calibrates the iteration count
per operation** to your Pi's speed, so results stay comparable across machines.
## Prerequisites
- **Raspberry Pi 5** (Cortex-A76, aarch64), ideally the 8GB model, with
**active cooling** so it doesn't thermal-throttle mid-run.
- **Raspberry Pi OS / Debian Trixie or newer** — system OpenSSL 3.5+ so PQ TLS
works without a source build.
- **Internet access** and **sudo**.
## Step 1 — Clone (public repo, no auth)
```sh
git clone <REPO_URL>
cd pq-bench-rpi5
```
## Step 2 — Build the toolchain
```sh
./setup/setup.sh all
```
Takes 5–15 min: installs dependencies and builds liboqs + oqs-provider. Run it
inside `tmux` so it survives an SSH disconnect.
## Step 3 — Run
```sh
sudo ./run.sh
```
`sudo` is needed to set the performance governor, pin cores, and read the
temperature. The run takes ~4–5 min, with no iteration counts to set.
Output lands in `results/<hostname>-<timestamp>.json`, stamped with full
provenance (Pi model, RAM, kernel, governor, thermal trace, library versions)
and an `is_baseline_grade` flag.
## Step 4 — View results
```sh
cd dashboard
python3 -m http.server 8765
# then open http://<pi-ip>:8765
```
The charts show KEM, signature, and TLS results with the classical X25519 /
Ed25519 baseline drawn as a reference line.
## Step 5 — Contribute (optional)
Share your `results/*.json` (open a PR or send it over). To merge results from
multiple machines:
```sh
python3 analyze/merge.py results/*.json -o dashboard/data/merged.json
```
The dashboard then shows every Pi side by side.
## What the results tell you
PQ is not so much *slower* as *bigger*. Lattice schemes (ML-KEM, ML-DSA) run
close to classical in speed but have much larger keys and signatures, while the
hash-based SLH-DSA (SPHINCS+) is an outlier in both signing time and signature
size. On TLS, the classical baseline fits in a single packet, while PQ and
hybrid handshakes grow past it and fragment.
## Notes and limitations
- Measures **liboqs** (C / assembly) implementations — a pure-Rust backend is a
separate, optional axis.
- Userspace PMU cycle counts are usually unavailable, so the primary metric is
**wall-clock time + ops/sec**.
- SNARK / STARK benchmarking is **out of scope** for this phase (`config.yaml`
reserves a hook for it).
- The candidate list lives in `config.yaml` — use the exact liboqs algorithm
names.

145
pq-bench-rpi5/analyze/merge.py Normal file
View File

@ -0,0 +1,145 @@
#!/usr/bin/env python3
"""merge.py — merge many results/<host>-<ts>.json files into one dataset the
dashboard (and PNG export) consume.
python3 analyze/merge.py results/*.json -o dashboard/data/merged.json
The merged file keeps every run as a separate record (so multiple machines /
repetitions can be compared) plus a flat index for quick charting. It never
collapses RPi5 baseline-grade runs together with non-baseline (e.g. macOS smoke)
runs — each record carries its own `is_baseline_grade` flag and host, and the
dashboard filters on it by default.
"""
from __future__ import annotations
import argparse
import glob
import json
import os
import sys
MERGED_SCHEMA = "1.0.0"
def load_runs(paths):
runs = []
for p in paths:
try:
with open(p) as f:
d = json.load(f)
d["_source_file"] = os.path.basename(p)
runs.append(d)
except (json.JSONDecodeError, OSError) as e:
print(f"[merge] skipping {p}: {e}", file=sys.stderr)
return runs
def run_id(run):
h = run.get("host", {})
return f"{h.get('hostname','?')}/{h.get('cpu_brand','?')}@{run.get('generated_utc','?')}"
def flatten(runs):
"""Produce flat per-(run, algorithm, operation) rows for easy charting."""
kem_rows, sig_rows, tls_rows = [], [], []
for run in runs:
rid = run_id(run)
host = run.get("host", {})
meta = {
"run_id": rid,
"hostname": host.get("hostname"),
"cpu_brand": host.get("cpu_brand"),
"is_rpi": host.get("is_rpi"),
"is_baseline_grade": run.get("is_baseline_grade"),
"source_file": run.get("_source_file"),
}
for k in run.get("kem", []):
if not k.get("enabled"):
continue
for op, st in (k.get("operations") or {}).items():
kem_rows.append({**meta,
"alg": k["alg"], "backend": k.get("backend"),
"classical": bool(k.get("classical")),
"nist_level": k.get("claimed_nist_level"),
"operation": op,
"median_ns": st.get("median"), "mad_ns": st.get("mad"),
"iqr_ns": st.get("iqr"), "min_ns": st.get("min"),
"stddev_ns": st.get("stddev"), "ops_per_sec": st.get("ops_per_sec"),
"sizes": k.get("sizes")})
for s in run.get("sig", []):
if not s.get("enabled"):
continue
for op, st in (s.get("operations") or {}).items():
sig_rows.append({**meta,
"alg": s["alg"], "backend": s.get("backend"),
"classical": bool(s.get("classical")),
"nist_level": s.get("claimed_nist_level"),
"operation": op,
"median_ns": st.get("median"), "mad_ns": st.get("mad"),
"iqr_ns": st.get("iqr"), "min_ns": st.get("min"),
"stddev_ns": st.get("stddev"), "ops_per_sec": st.get("ops_per_sec"),
"sizes": s.get("sizes")})
tls = run.get("tls") or {}
for cell in (tls.get("matrix") or []):
if not cell.get("enabled"):
continue
tls_rows.append({**meta,
"label": cell.get("label"), "group": cell.get("group"),
"is_baseline_pair": cell.get("label") == (tls.get("baseline") or {}).get("label"),
"handshakes_per_sec": cell.get("handshakes_per_sec"),
"median_ns": (cell.get("handshake_latency_ns") or {}).get("median"),
"bytes_total": (cell.get("bytes_on_wire") or {}).get("total"),
"client_hello_bytes": cell.get("client_hello_bytes"),
"client_hello_fragmented": cell.get("client_hello_fragmented")})
return kem_rows, sig_rows, tls_rows
def main():
ap = argparse.ArgumentParser()
ap.add_argument("inputs", nargs="+", help="results JSON files or globs")
ap.add_argument("-o", "--out", default="dashboard/data/merged.json")
args = ap.parse_args()
paths = []
for pat in args.inputs:
paths.extend(sorted(glob.glob(pat)) if any(c in pat for c in "*?[") else [pat])
paths = [p for p in paths if os.path.isfile(p)]
if not paths:
sys.exit("no input files matched")
runs = load_runs(paths)
kem_rows, sig_rows, tls_rows = flatten(runs)
merged = {
"merged_schema": MERGED_SCHEMA,
"n_runs": len(runs),
"runs": [{
"run_id": run_id(r),
"host": r.get("host"),
"is_baseline_grade": r.get("is_baseline_grade"),
"baseline_grade_reasons": r.get("baseline_grade_reasons", []),
"toolchain": r.get("toolchain"),
"cpu_features": r.get("cpu_features"),
"run": r.get("run"),
"thermal_summary": {
"temp_c": (r.get("thermal_trace") or {}).get("temp_c"),
"throttling_detected": (r.get("thermal_trace") or {}).get("throttling_detected"),
},
"generated_utc": r.get("generated_utc"),
"source_file": r.get("_source_file"),
} for r in runs],
"kem": kem_rows,
"sig": sig_rows,
"tls": tls_rows,
}
os.makedirs(os.path.dirname(args.out) or ".", exist_ok=True)
with open(args.out, "w") as f:
json.dump(merged, f, indent=2)
n_base = sum(1 for r in runs if r.get("is_baseline_grade"))
print(f"merged {len(runs)} run(s) -> {args.out} "
f"({n_base} baseline-grade, {len(runs)-n_base} smoke/other)")
print(f" kem rows: {len(kem_rows)} sig rows: {len(sig_rows)} tls rows: {len(tls_rows)}")
if __name__ == "__main__":
main()

162
pq-bench-rpi5/analyze/plot.py Normal file
View File

@ -0,0 +1,162 @@
#!/usr/bin/env python3
"""plot.py — export publication-ready PNGs from a merged dataset.
python3 analyze/plot.py dashboard/data/merged.json -o analyze/png
matplotlib is an OPTIONAL dependency. If it is not installed this prints a clear
hint and exits 0 (the HTML dashboard remains the primary, dependency-free view).
By default it plots only baseline-grade (RPi5) runs so paper figures are never
polluted with macOS smoke data; pass --include-smoke to override. The classical
Logos baseline (X25519 / Ed25519) is drawn as a reference line on every chart.
"""
from __future__ import annotations
import argparse
import json
import os
import sys
try:
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
except ImportError:
print("matplotlib not installed — skipping PNG export.\n"
" enable it in a project venv (keeps your system python clean):\n"
" python3 -m venv analyze/.venv\n"
" analyze/.venv/bin/pip install -r analyze/requirements.txt\n"
" analyze/.venv/bin/python analyze/plot.py dashboard/data/merged.json\n"
" (the HTML dashboard works without it)", file=sys.stderr)
sys.exit(0)
def median_ms(ns):
return (ns or 0) / 1e6
def pick_runs(merged, include_smoke):
ids = set()
for r in merged.get("runs", []):
if include_smoke or r.get("is_baseline_grade"):
ids.add(r["run_id"])
return ids
def grouped_bar_by_level(rows, op, title, outpath):
"""Grouped bars of median latency per algorithm, grouped by NIST level."""
data = [r for r in rows if r["operation"] == op]
if not data:
return
# one bar per algorithm; baseline highlighted
data.sort(key=lambda r: (r.get("nist_level") or 0, r["median_ns"]))
labels = [r["alg"] for r in data]
vals = [median_ms(r["median_ns"]) for r in data]
colors = ["#888" if r.get("classical") else "#3b6" for r in data]
fig, ax = plt.subplots(figsize=(max(6, len(labels) * 0.55), 4))
ax.bar(range(len(labels)), vals, color=colors)
# baseline reference line
base = next((r for r in data if r.get("classical")), None)
if base:
ax.axhline(median_ms(base["median_ns"]), color="#c33", ls="--", lw=1,
label=f"classical baseline ({base['alg']})")
ax.legend(fontsize=8)
ax.set_xticks(range(len(labels)))
ax.set_xticklabels(labels, rotation=45, ha="right", fontsize=8)
ax.set_ylabel("median latency (ms)")
ax.set_title(title)
ax.grid(axis="y", alpha=0.3)
fig.tight_layout()
fig.savefig(outpath, dpi=150)
plt.close(fig)
print("wrote", outpath)
def size_speed_scatter(rows, op, size_key, title, outpath):
data = [r for r in rows if r["operation"] == op and r.get("sizes")]
pts = []
for r in data:
sz = (r["sizes"] or {}).get(size_key)
if sz:
pts.append((sz, median_ms(r["median_ns"]), r["alg"], r.get("classical")))
if not pts:
return
fig, ax = plt.subplots(figsize=(7, 5))
for sz, lat, alg, classical in pts:
ax.scatter(sz, lat, c="#c33" if classical else "#3b6",
s=60, edgecolors="k", linewidths=0.4, zorder=3)
ax.annotate(alg, (sz, lat), fontsize=7, xytext=(4, 3),
textcoords="offset points")
ax.set_xlabel(f"{size_key} size (bytes)")
ax.set_ylabel("median latency (ms)")
ax.set_title(title)
ax.grid(alpha=0.3)
fig.tight_layout()
fig.savefig(outpath, dpi=150)
plt.close(fig)
print("wrote", outpath)
def tls_bar(tls_rows, outpath):
if not tls_rows:
return
tls_rows = sorted(tls_rows, key=lambda r: -(r.get("handshakes_per_sec") or 0))
labels = [r["label"] for r in tls_rows]
vals = [r.get("handshakes_per_sec") or 0 for r in tls_rows]
colors = ["#c33" if r.get("is_baseline_pair") else "#36c" for r in tls_rows]
fig, ax = plt.subplots(figsize=(max(6, len(labels) * 0.5), 4.5))
ax.barh(range(len(labels)), vals, color=colors)
base = next((r for r in tls_rows if r.get("is_baseline_pair")), None)
if base:
ax.axvline(base.get("handshakes_per_sec") or 0, color="#c33", ls="--", lw=1,
label=f"classical baseline ({base['label']})")
ax.legend(fontsize=8)
ax.set_yticks(range(len(labels)))
ax.set_yticklabels(labels, fontsize=7)
ax.invert_yaxis()
ax.set_xlabel("handshakes / sec")
ax.set_title("TLS 1.3 handshake throughput (higher is better)")
ax.grid(axis="x", alpha=0.3)
fig.tight_layout()
fig.savefig(outpath, dpi=150)
plt.close(fig)
print("wrote", outpath)
def main():
ap = argparse.ArgumentParser()
ap.add_argument("merged")
ap.add_argument("-o", "--outdir", default="analyze/png")
ap.add_argument("--include-smoke", action="store_true")
args = ap.parse_args()
with open(args.merged) as f:
merged = json.load(f)
ids = pick_runs(merged, args.include_smoke)
if not ids:
print("no baseline-grade runs to plot (use --include-smoke for macOS/dev data)",
file=sys.stderr)
return
def keep(rows):
return [r for r in rows if r["run_id"] in ids]
kem, sig, tls = keep(merged["kem"]), keep(merged["sig"]), keep(merged["tls"])
os.makedirs(args.outdir, exist_ok=True)
O = args.outdir
grouped_bar_by_level(kem, "keygen", "KEM keygen latency by algorithm", f"{O}/kem_keygen.png")
grouped_bar_by_level(kem, "encaps", "KEM encapsulation latency", f"{O}/kem_encaps.png")
grouped_bar_by_level(kem, "decaps", "KEM decapsulation latency", f"{O}/kem_decaps.png")
grouped_bar_by_level(sig, "sign", "Signature signing latency", f"{O}/sig_sign.png")
grouped_bar_by_level(sig, "verify", "Signature verification latency", f"{O}/sig_verify.png")
size_speed_scatter(kem, "encaps", "public_key",
"KEM: public-key size vs encaps latency", f"{O}/kem_size_speed.png")
size_speed_scatter(sig, "sign", "signature",
"Signature: signature size vs sign latency", f"{O}/sig_size_speed.png")
tls_bar(tls, f"{O}/tls_throughput.png")
print(f"PNGs in {O}/ ({'incl. smoke' if args.include_smoke else 'baseline-grade only'})")
if __name__ == "__main__":
main()

1
pq-bench-rpi5/analyze/requirements.txt Normal file
View File

@ -0,0 +1 @@
matplotlib>=3.7

26
pq-bench-rpi5/bench/kem_sig/Makefile Normal file
View File

@ -0,0 +1,26 @@
# Build the primitive KEM/sig harness against the pinned liboqs + system OpenSSL.
#
# Paths/flags default to the vendored toolchain but are overridable so run.sh can
# pass the exact values from setup/versions.lock (single source of truth):
# make LIBOQS_PREFIX=... OPENSSL_PREFIX=... BENCH_CFLAGS="-O3 -mcpu=cortex-a76"
ROOT := $(abspath $(dir $(lastword $(MAKEFILE_LIST)))../..)
LIBOQS_PREFIX ?= $(ROOT)/vendor/install
# OpenSSL: prefer Homebrew openssl@3 on macOS, else system.
OPENSSL_PREFIX ?= $(shell brew --prefix openssl@3 2>/dev/null || echo /usr)
BENCH_CFLAGS ?= -O3
CC ?= cc
CFLAGS := $(BENCH_CFLAGS) -std=c11 -Wall -Wextra \
-I$(LIBOQS_PREFIX)/include -I$(OPENSSL_PREFIX)/include
LDFLAGS := -L$(LIBOQS_PREFIX)/lib -L$(OPENSSL_PREFIX)/lib \
-Wl,-rpath,$(LIBOQS_PREFIX)/lib -Wl,-rpath,$(OPENSSL_PREFIX)/lib
LDLIBS := -loqs -lcrypto -lm
bench_pq: bench_pq.c
$(CC) $(CFLAGS) -o $@ $< $(LDFLAGS) $(LDLIBS)
clean:
rm -f bench_pq
.PHONY: clean

656
pq-bench-rpi5/bench/kem_sig/bench_pq.c Normal file
View File

@ -0,0 +1,656 @@
/* ===========================================================================
* bench_pq.c — primitive-level KEM / signature benchmark harness.
*
* One algorithm per invocation (fresh process keeps cache state clean):
* bench_pq --kind kem --alg ML-KEM-768 --warmup 1000 --iters 10000 --reps 5
* bench_pq --kind sig --alg ML-DSA-65 ...
*
* Emits a single JSON object describing the algorithm to stdout. The orchestrator
* (run.sh / assemble.py) wraps these with environment metadata.
*
* Two backends, selected by algorithm name:
* - liboqs : all PQ candidates (ML-KEM, ML-DSA, Falcon, SLH-DSA, ...)
* - OpenSSL EVP : the classical Logos baselines X25519 (KEM-analog) and
* Ed25519 (signature), which liboqs does not implement.
* This lets the classical reference be drawn on the same primitive charts.
*
* Metrics per operation: full per-iteration wall-clock nanosecond distribution
* -> median, MAD, IQR, min, max, mean, stddev, ops/sec, plus per-repetition
* medians. Optional userspace PMU cycle counts when available. Heap high-water
* via mallinfo2 on glibc (the RPi5 target); honestly reported unavailable
* elsewhere (e.g. the macOS smoke box).
* ===========================================================================*/
#define _POSIX_C_SOURCE 200809L
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
#include <math.h>
#include <time.h>
#include <setjmp.h>
#include <signal.h>
#include <oqs/oqs.h>
#include <openssl/evp.h>
#include <openssl/err.h>
#if defined(__linux__) && defined(__GLIBC__)
#include <malloc.h>
#define HAVE_MALLINFO2 1
#endif
/* ---- timing ------------------------------------------------------------- */
static inline uint64_t now_ns(void) {
struct timespec ts;
clock_gettime(CLOCK_MONOTONIC, &ts);
return (uint64_t)ts.tv_sec * 1000000000ull + (uint64_t)ts.tv_nsec;
}
/* ---- userspace PMU cycle counter probe (aarch64) ------------------------ */
static sigjmp_buf g_sigill_jmp;
static volatile sig_atomic_t g_pmu_ok = 0;
static void sigill_handler(int sig) { (void)sig; siglongjmp(g_sigill_jmp, 1); }
static inline uint64_t read_cycles(void) {
#if defined(__aarch64__)
uint64_t v;
__asm__ volatile("mrs %0, pmccntr_el0" : "=r"(v));
return v;
#else
return 0;
#endif
}
/* Returns 1 and a reason string if userspace cycle counting works. */
static int probe_pmu(const char **reason) {
#if defined(__aarch64__)
struct sigaction sa, old;
memset(&sa, 0, sizeof sa);
sa.sa_handler = sigill_handler;
sigaction(SIGILL, &sa, &old);
if (sigsetjmp(g_sigill_jmp, 1) == 0) {
(void)read_cycles();
g_pmu_ok = 1;
} else {
g_pmu_ok = 0;
}
sigaction(SIGILL, &old, NULL);
if (g_pmu_ok) { *reason = "PMCCNTR_EL0 readable from userspace"; return 1; }
*reason = "PMCCNTR_EL0 traps (kernel module not loaded; needs e.g. enable_arm_pmu)";
return 0;
#else
*reason = "not aarch64";
return 0;
#endif
}
/* ---- statistics --------------------------------------------------------- */
typedef struct {
double median, mad, iqr, q1, q3, min, max, mean, stddev;
double ops_per_sec;
uint64_t n;
} stats_t;
static int cmp_u64(const void *a, const void *b) {
uint64_t x = *(const uint64_t *)a, y = *(const uint64_t *)b;
return (x > y) - (x < y);
}
/* percentile on an already-sorted array (linear interpolation) */
static double pct_sorted(const uint64_t *s, uint64_t n, double p) {
if (n == 0) return 0;
if (n == 1) return (double)s[0];
double idx = p * (double)(n - 1);
uint64_t lo = (uint64_t)idx;
double frac = idx - (double)lo;
if (lo + 1 >= n) return (double)s[n - 1];
return (double)s[lo] + frac * ((double)s[lo + 1] - (double)s[lo]);
}
static stats_t compute_stats(uint64_t *samples, uint64_t n) {
stats_t st; memset(&st, 0, sizeof st);
st.n = n;
if (n == 0) return st;
qsort(samples, n, sizeof(uint64_t), cmp_u64);
st.min = (double)samples[0];
st.max = (double)samples[n - 1];
st.median = pct_sorted(samples, n, 0.5);
st.q1 = pct_sorted(samples, n, 0.25);
st.q3 = pct_sorted(samples, n, 0.75);
st.iqr = st.q3 - st.q1;
double sum = 0;
for (uint64_t i = 0; i < n; i++) sum += (double)samples[i];
st.mean = sum / (double)n;
double ss = 0;
for (uint64_t i = 0; i < n; i++) {
double d = (double)samples[i] - st.mean;
ss += d * d;
}
st.stddev = (n > 1) ? sqrt(ss / (double)(n - 1)) : 0;
/* MAD = median(|x - median|); needs a second sorted buffer */
uint64_t *dev = malloc(n * sizeof(uint64_t));
if (dev) {
for (uint64_t i = 0; i < n; i++) {
double d = (double)samples[i] - st.median;
dev[i] = (uint64_t)(d < 0 ? -d : d);
}
qsort(dev, n, sizeof(uint64_t), cmp_u64);
st.mad = pct_sorted(dev, n, 0.5);
free(dev);
}
st.ops_per_sec = st.median > 0 ? 1e9 / st.median : 0;
return st;
}
/* ---- measurement configuration + generic loop --------------------------- */
/* A closure-ish: the caller provides a function pointer that runs one op. */
typedef int (*op_fn)(void *ctx);
/* How each op is sized. Two modes, decided per invocation:
* - fixed-count (fixed_iters > 0): exactly fixed_iters timed iters per rep,
* with the configured fixed-mode warmup. This is the explicit
* `--iters N` override / fixed-count fallback.
* - auto-calibrate(fixed_iters == 0): each op is timed long enough to put
* ~target_ns of aggregate timed work on it, so a 18 us keygen
* and a 750 ms SLH-DSA sign both yield a stable median. The
* chosen count is clamped to [min_samples, max_iters]:
* fast ops hit max_iters, slow ops hit min_samples.
* Auto mode estimates per-op cost with a short doubling calibration (which also
* serves as cache warmup), then derives the timed count + a per-rep re-warm. */
typedef struct {
uint64_t fixed_iters; /* >0 => fixed-count mode; 0 => auto-calibrate */
uint64_t target_ns; /* auto: per-op aggregate timed-work target */
uint64_t min_samples; /* auto: floor on timed iters per rep */
uint64_t max_iters; /* auto: ceiling on timed iters per rep */
uint64_t warmup; /* fixed-mode warmup count per rep */
uint64_t reps;
} bench_cfg;
typedef struct {
uint64_t *all; /* all samples across reps */
uint64_t all_n;
double per_rep_median[64];
int n_rep_med;
uint64_t timed_iters; /* timed iters per rep actually used */
uint64_t warmup_iters; /* warmup iters per rep actually used */
uint64_t reps;
int calibrated; /* 1 if auto-calibrated, 0 if fixed-count */
double est_ns; /* per-op cost estimate from calibration (auto) */
} measure_out;
static void print_stats_json(FILE *f, const char *name, stats_t st,
const measure_out *m) {
fprintf(f, "\"%s\":{", name);
fprintf(f, "\"unit\":\"ns\",\"warmup_iters\":%llu,\"timed_iters\":%llu,\"repetitions\":%llu,",
(unsigned long long)m->warmup_iters, (unsigned long long)m->timed_iters,
(unsigned long long)m->reps);
fprintf(f, "\"calibrated\":%s,", m->calibrated ? "true" : "false");
if (m->calibrated)
fprintf(f, "\"calib_est_ns\":%.2f,", m->est_ns);
fprintf(f, "\"samples\":%llu,", (unsigned long long)st.n);
fprintf(f, "\"median\":%.2f,\"mad\":%.2f,\"iqr\":%.2f,\"q1\":%.2f,\"q3\":%.2f,",
st.median, st.mad, st.iqr, st.q1, st.q3);
fprintf(f, "\"min\":%.2f,\"max\":%.2f,\"mean\":%.2f,\"stddev\":%.2f,",
st.min, st.max, st.mean, st.stddev);
fprintf(f, "\"ops_per_sec\":%.2f,", st.ops_per_sec);
fprintf(f, "\"per_rep_median\":[");
for (int i = 0; i < m->n_rep_med; i++)
fprintf(f, "%s%.2f", i ? "," : "", m->per_rep_median[i]);
fprintf(f, "]}");
}
/* Pick the timed-iter count (and per-rep warmup) for one op under cfg.
* In auto mode this runs a doubling calibration loop on fn (which warms caches)
* to estimate per-op cost, then solves for the count that hits target_ns.
* Returns 0 on success, -2 if the op ever fails during calibration. */
static int calibrate_op(op_fn fn, void *ctx, const bench_cfg *cfg,
uint64_t *timed_out, uint64_t *warm_out,
double *est_ns_out, int *calibrated_out) {
if (cfg->fixed_iters > 0) {
*timed_out = cfg->fixed_iters;
*warm_out = cfg->warmup;
*est_ns_out = 0.0;
*calibrated_out = 0;
return 0;
}
/* doubling calibration: run batches 1,2,4,... until ~CALIB_BUDGET elapses,
* capped at max_iters so a sub-microsecond op can't spin forever. */
const uint64_t CALIB_BUDGET_NS = 30ull * 1000 * 1000; /* 30 ms */
uint64_t cops = 0, cel = 0, batch = 1;
while (cel < CALIB_BUDGET_NS && cops < cfg->max_iters) {
uint64_t t0 = now_ns();
for (uint64_t i = 0; i < batch; i++)
if (fn(ctx) != 0) return -2;
cel += now_ns() - t0;
cops += batch;
batch *= 2;
}
double est_ns = cops ? (double)cel / (double)cops : 1.0;
if (est_ns < 1.0) est_ns = 1.0;
double want = (double)cfg->target_ns / est_ns;
uint64_t n = (uint64_t)(want + 0.5);
if (n < cfg->min_samples) n = cfg->min_samples; /* slow ops floor here */
if (n > cfg->max_iters) n = cfg->max_iters; /* fast ops ceil here */
/* per-rep re-warm ~= 20% of the timed budget, at least 1, capped. */
double w = ((double)cfg->target_ns * 0.2) / est_ns;
uint64_t warm = (uint64_t)(w + 0.5);
if (warm < 1) warm = 1;
if (warm > cfg->max_iters) warm = cfg->max_iters;
*timed_out = n;
*warm_out = warm;
*est_ns_out = est_ns;
*calibrated_out = 1;
return 0;
}
/* Returns 0 on success. Fills out with samples. Re-warms before each rep. */
static int measure_op(op_fn fn, void *ctx, const bench_cfg *cfg, measure_out *out) {
uint64_t iters, warmup;
double est_ns; int calibrated;
if (calibrate_op(fn, ctx, cfg, &iters, &warmup, &est_ns, &calibrated) != 0)
return -2;
out->all = malloc(iters * cfg->reps * sizeof(uint64_t));
if (!out->all) return -1;
out->all_n = 0;
out->n_rep_med = 0;
out->timed_iters = iters;
out->warmup_iters = warmup;
out->reps = cfg->reps;
out->calibrated = calibrated;
out->est_ns = est_ns;
uint64_t *rep_buf = malloc(iters * sizeof(uint64_t));
if (!rep_buf) { free(out->all); return -1; }
for (uint64_t r = 0; r < cfg->reps; r++) {
for (uint64_t i = 0; i < warmup; i++)
if (fn(ctx) != 0) { free(rep_buf); free(out->all); return -2; }
for (uint64_t i = 0; i < iters; i++) {
uint64_t t0 = now_ns();
int rc = fn(ctx);
uint64_t dt = now_ns() - t0;
if (rc != 0) { free(rep_buf); free(out->all); return -2; }
rep_buf[i] = dt;
out->all[out->all_n++] = dt;
}
/* per-rep median (sorts a copy of this rep's slice) */
uint64_t *copy = malloc(iters * sizeof(uint64_t));
if (copy) {
memcpy(copy, rep_buf, iters * sizeof(uint64_t));
qsort(copy, iters, sizeof(uint64_t), cmp_u64);
if (out->n_rep_med < 64)
out->per_rep_median[out->n_rep_med++] = pct_sorted(copy, iters, 0.5);
free(copy);
}
}
free(rep_buf);
return 0;
}
/* ---- anti-DCE sink + fatal helpers -------------------------------------- */
/* File-scope volatile: the compiler must materialize every store, so it cannot
* dead-code-eliminate the crypto outputs we feed into it. Every timed op sinks
* one output byte here. */
static volatile uint64_t g_sink = 0;
/* Abort the whole run: a broken build must NEVER silently emit timing numbers. */
static void die(const char *alg, const char *what) {
fprintf(stderr, "[bench_pq] FATAL: %s: %s — aborting so no numbers are emitted\n",
alg, what);
exit(3);
}
/* measure_op wrapper: if any timed op ever returns failure (e.g. a verify that
* stopped succeeding), abort instead of reading freed/partial buffers. */
static void must_measure(const char *alg, const char *op, op_fn fn, void *ctx,
const bench_cfg *cfg, measure_out *out) {
if (measure_op(fn, ctx, cfg, out) != 0)
die(alg, op);
}
#define MSGLEN 32
/* ======================= liboqs KEM ====================================== */
/* Timed ops consume canonical, pre-validated inputs; keygen/encaps write to
* scratch buffers so they never clobber the matched (pk,sk,ct) that the other
* timed ops depend on. Each op sinks an output byte into g_sink. */
typedef struct {
OQS_KEM *kem;
uint8_t *pk, *sk; /* canonical matched keypair (encaps/decaps inputs) */
uint8_t *ct; /* canonical ciphertext (decaps input) */
uint8_t *pk_s, *sk_s; /* scratch: keygen outputs */
uint8_t *ct_s, *ss_s; /* scratch: encaps outputs */
uint8_t *ss_d; /* scratch: decaps output */
} kem_ctx;
static int kem_keygen(void *c){ kem_ctx*x=c;
if (OQS_KEM_keypair(x->kem, x->pk_s, x->sk_s) != OQS_SUCCESS) return 1;
g_sink += x->pk_s[0]; return 0; }
static int kem_encaps(void *c){ kem_ctx*x=c;
if (OQS_KEM_encaps(x->kem, x->ct_s, x->ss_s, x->pk) != OQS_SUCCESS) return 1;
g_sink += (uint64_t)x->ss_s[0] ^ x->ct_s[0]; return 0; }
static int kem_decaps(void *c){ kem_ctx*x=c;
if (OQS_KEM_decaps(x->kem, x->ss_d, x->ct, x->sk) != OQS_SUCCESS) return 1;
g_sink += x->ss_d[0]; return 0; }
static int run_kem(const char *alg, const bench_cfg *cfg) {
OQS_KEM *kem = OQS_KEM_new(alg);
if (!kem) {
printf("{\"alg\":\"%s\",\"kind\":\"kem\",\"backend\":\"liboqs\",\"enabled\":false,"
"\"reason\":\"not enabled in this liboqs build\"}\n", alg);
return 0;
}
kem_ctx x; memset(&x, 0, sizeof x); x.kem = kem;
x.pk = malloc(kem->length_public_key); x.sk = malloc(kem->length_secret_key);
x.ct = malloc(kem->length_ciphertext);
x.pk_s = malloc(kem->length_public_key); x.sk_s = malloc(kem->length_secret_key);
x.ct_s = malloc(kem->length_ciphertext);
x.ss_s = malloc(kem->length_shared_secret);
x.ss_d = malloc(kem->length_shared_secret);
if (!x.pk||!x.sk||!x.ct||!x.pk_s||!x.sk_s||!x.ct_s||!x.ss_s||!x.ss_d) die(alg,"out of memory");
/* ---- correctness check (ONCE, outside the timed loop) ----
* keygen -> encaps -> decaps, then assert the shared secrets match. */
if (OQS_KEM_keypair(kem, x.pk, x.sk) != OQS_SUCCESS) die(alg, "keygen failed");
if (OQS_KEM_encaps(kem, x.ct, x.ss_s, x.pk) != OQS_SUCCESS) die(alg, "encaps failed");
if (OQS_KEM_decaps(kem, x.ss_d, x.ct, x.sk) != OQS_SUCCESS) die(alg, "decaps failed");
if (memcmp(x.ss_s, x.ss_d, kem->length_shared_secret) != 0)
die(alg, "KEM shared-secret mismatch (ss_encaps != ss_decaps)");
/* timed phases run on the canonical, validated (pk,sk,ct); any op failure
* during timing aborts via must_measure. */
measure_out kg={0}, en={0}, de={0};
must_measure(alg,"keygen",kem_keygen,&x,cfg,&kg);
must_measure(alg,"encaps",kem_encaps,&x,cfg,&en);
must_measure(alg,"decaps",kem_decaps,&x,cfg,&de);
printf("{\"alg\":\"%s\",\"kind\":\"kem\",\"backend\":\"liboqs\",\"enabled\":true,", alg);
printf("\"claimed_nist_level\":%d,", kem->claimed_nist_level);
printf("\"sizes\":{\"public_key\":%zu,\"secret_key\":%zu,\"ciphertext\":%zu,\"shared_secret\":%zu},",
kem->length_public_key, kem->length_secret_key, kem->length_ciphertext, kem->length_shared_secret);
printf("\"operations\":{");
stats_t s;
s=compute_stats(kg.all,kg.all_n); print_stats_json(stdout,"keygen",s,&kg); printf(",");
s=compute_stats(en.all,en.all_n); print_stats_json(stdout,"encaps",s,&en); printf(",");
s=compute_stats(de.all,de.all_n); print_stats_json(stdout,"decaps",s,&de);
printf("}}\n");
free(kg.all); free(en.all); free(de.all);
free(x.pk); free(x.sk); free(x.ct);
free(x.pk_s); free(x.sk_s); free(x.ct_s); free(x.ss_s); free(x.ss_d);
OQS_KEM_free(kem);
return 0;
}
/* ======================= liboqs SIG ====================================== */
/* sign writes a scratch signature; verify reads the canonical (sg,sglen) over
* the canonical msg with the canonical pk — all pre-validated. */
typedef struct {
OQS_SIG *sig;
uint8_t *pk, *sk; /* canonical keypair (sign/verify inputs) */
uint8_t *msg; /* canonical message */
uint8_t *sg; size_t sglen; /* canonical signature (verify input) */
uint8_t *pk_s, *sk_s; /* scratch: keygen outputs */
uint8_t *sg_s; size_t sg_s_len;/* scratch: sign output */
} sig_ctx;
static int sig_keygen(void *c){ sig_ctx*x=c;
if (OQS_SIG_keypair(x->sig, x->pk_s, x->sk_s) != OQS_SUCCESS) return 1;
g_sink += x->pk_s[0]; return 0; }
static int sig_sign(void *c){ sig_ctx*x=c;
x->sg_s_len = x->sig->length_signature;
if (OQS_SIG_sign(x->sig, x->sg_s, &x->sg_s_len, x->msg, MSGLEN, x->sk) != OQS_SUCCESS) return 1;
g_sink += x->sg_s[0]; return 0; }
static int sig_verify(void *c){ sig_ctx*x=c;
if (OQS_SIG_verify(x->sig, x->msg, MSGLEN, x->sg, x->sglen, x->pk) != OQS_SUCCESS) return 1;
g_sink += 1; return 0; }
static int run_sig(const char *alg, const bench_cfg *cfg) {
OQS_SIG *sig = OQS_SIG_new(alg);
if (!sig) {
printf("{\"alg\":\"%s\",\"kind\":\"sig\",\"backend\":\"liboqs\",\"enabled\":false,"
"\"reason\":\"not enabled in this liboqs build\"}\n", alg);
return 0;
}
sig_ctx x; memset(&x,0,sizeof x); x.sig=sig;
x.pk = malloc(sig->length_public_key); x.sk = malloc(sig->length_secret_key);
x.msg = malloc(MSGLEN);
x.sg = malloc(sig->length_signature);
x.pk_s = malloc(sig->length_public_key); x.sk_s = malloc(sig->length_secret_key);
x.sg_s = malloc(sig->length_signature);
if (!x.pk||!x.sk||!x.msg||!x.sg||!x.pk_s||!x.sk_s||!x.sg_s) die(alg,"out of memory");
memset(x.msg, 0xA5, MSGLEN);
/* ---- correctness check (ONCE, outside the timed loop) ----
* keygen -> sign -> verify; the verify MUST succeed on a valid signature. */
if (OQS_SIG_keypair(sig, x.pk, x.sk) != OQS_SUCCESS) die(alg, "keygen failed");
x.sglen = sig->length_signature;
if (OQS_SIG_sign(sig, x.sg, &x.sglen, x.msg, MSGLEN, x.sk) != OQS_SUCCESS) die(alg, "sign failed");
if (OQS_SIG_verify(sig, x.msg, MSGLEN, x.sg, x.sglen, x.pk) != OQS_SUCCESS)
die(alg, "signature verify failed on a valid signature (broken build)");
measure_out kg={0}, sg={0}, vf={0};
must_measure(alg,"keygen",sig_keygen,&x,cfg,&kg);
must_measure(alg,"sign", sig_sign, &x,cfg,&sg);
must_measure(alg,"verify",sig_verify,&x,cfg,&vf);
printf("{\"alg\":\"%s\",\"kind\":\"sig\",\"backend\":\"liboqs\",\"enabled\":true,", alg);
printf("\"claimed_nist_level\":%d,", sig->claimed_nist_level);
printf("\"sizes\":{\"public_key\":%zu,\"secret_key\":%zu,\"signature\":%zu},",
sig->length_public_key, sig->length_secret_key, sig->length_signature);
printf("\"operations\":{");
stats_t s;
s=compute_stats(kg.all,kg.all_n); print_stats_json(stdout,"keygen",s,&kg); printf(",");
s=compute_stats(sg.all,sg.all_n); print_stats_json(stdout,"sign",s,&sg); printf(",");
s=compute_stats(vf.all,vf.all_n); print_stats_json(stdout,"verify",s,&vf);
printf("}}\n");
free(kg.all); free(sg.all); free(vf.all);
free(x.pk); free(x.sk); free(x.msg); free(x.sg);
free(x.pk_s); free(x.sk_s); free(x.sg_s);
OQS_SIG_free(sig);
return 0;
}
/* ======================= OpenSSL classical baselines ===================== */
/* X25519 as a KEM-analog: keygen + ECDH derive (one shared-secret derivation). */
typedef struct { EVP_PKEY *self; EVP_PKEY *peer; } x25519_ctx;
static int x25519_keygen(void *c){
x25519_ctx*x=c;
if (x->self) { EVP_PKEY_free(x->self); x->self=NULL; }
EVP_PKEY_CTX *p = EVP_PKEY_CTX_new_id(EVP_PKEY_X25519,NULL);
if(!p) return 1;
int ok = EVP_PKEY_keygen_init(p)>0 && EVP_PKEY_keygen(p,&x->self)>0;
EVP_PKEY_CTX_free(p);
return ok?0:1;
}
/* derive shared secret a·b into out[32]; returns 1 on success */
static int x25519_derive_into(EVP_PKEY *a, EVP_PKEY *b, unsigned char out[32]){
EVP_PKEY_CTX *p = EVP_PKEY_CTX_new(a,NULL);
if(!p) return 0;
size_t slen=32;
int ok = EVP_PKEY_derive_init(p)>0 &&
EVP_PKEY_derive_set_peer(p,b)>0 &&
EVP_PKEY_derive(p,out,&slen)>0;
EVP_PKEY_CTX_free(p);
return ok;
}
static int x25519_derive(void *c){
x25519_ctx*x=c;
unsigned char secret[32];
if(!x25519_derive_into(x->self,x->peer,secret)) return 1;
g_sink += secret[0]; /* sink the derived shared secret */
return 0;
}
static int run_x25519(const bench_cfg *cfg) {
x25519_ctx x={0};
if (x25519_keygen(&x) != 0) die("X25519","keygen failed"); /* self */
/* a fixed peer key for derive */
EVP_PKEY_CTX *p = EVP_PKEY_CTX_new_id(EVP_PKEY_X25519,NULL);
if(!p || EVP_PKEY_keygen_init(p)<=0 || EVP_PKEY_keygen(p,&x.peer)<=0) die("X25519","peer keygen failed");
EVP_PKEY_CTX_free(p);
/* ---- correctness check (ONCE, outside timing): ECDH must agree ---- */
{
unsigned char sa[32], sb[32];
if (!x25519_derive_into(x.self, x.peer, sa)) die("X25519","derive(self,peer) failed");
if (!x25519_derive_into(x.peer, x.self, sb)) die("X25519","derive(peer,self) failed");
if (memcmp(sa, sb, 32) != 0) die("X25519","ECDH shared-secret mismatch");
}
measure_out kg={0}, dv={0};
must_measure("X25519","keygen",x25519_keygen,&x,cfg,&kg);
/* keygen frees+replaces self each call; re-make a stable self for derive */
if (x25519_keygen(&x) != 0) die("X25519","keygen failed");
must_measure("X25519","derive",x25519_derive,&x,cfg,&dv);
printf("{\"alg\":\"X25519\",\"kind\":\"kem\",\"backend\":\"openssl\",\"classical\":true,\"enabled\":true,");
printf("\"claimed_nist_level\":1,");
printf("\"sizes\":{\"public_key\":32,\"secret_key\":32,\"ciphertext\":null,\"shared_secret\":32},");
printf("\"operations\":{");
stats_t s;
s=compute_stats(kg.all,kg.all_n); print_stats_json(stdout,"keygen",s,&kg); printf(",");
s=compute_stats(dv.all,dv.all_n); print_stats_json(stdout,"derive",s,&dv);
printf("}}\n");
free(kg.all); free(dv.all);
if(x.self)EVP_PKEY_free(x.self); if(x.peer)EVP_PKEY_free(x.peer);
return 0;
}
/* Ed25519 signature baseline. */
typedef struct { EVP_PKEY *key; unsigned char msg[32]; unsigned char sig[64]; size_t siglen; } ed_ctx;
static int ed_keygen(void *c){
ed_ctx*x=c;
if(x->key){EVP_PKEY_free(x->key);x->key=NULL;}
EVP_PKEY_CTX *p=EVP_PKEY_CTX_new_id(EVP_PKEY_ED25519,NULL);
if(!p)return 1;
int ok=EVP_PKEY_keygen_init(p)>0 && EVP_PKEY_keygen(p,&x->key)>0;
EVP_PKEY_CTX_free(p);
return ok?0:1;
}
static int ed_sign(void *c){
ed_ctx*x=c;
EVP_MD_CTX *m=EVP_MD_CTX_new(); if(!m)return 1;
x->siglen=sizeof x->sig;
int ok = EVP_DigestSignInit(m,NULL,NULL,NULL,x->key)>0 &&
EVP_DigestSign(m,x->sig,&x->siglen,x->msg,sizeof x->msg)>0;
EVP_MD_CTX_free(m);
if(!ok) return 1;
g_sink += x->sig[0]; /* sink the signature */
return 0;
}
static int ed_verify(void *c){
ed_ctx*x=c;
EVP_MD_CTX *m=EVP_MD_CTX_new(); if(!m)return 1;
int ok = EVP_DigestVerifyInit(m,NULL,NULL,NULL,x->key)>0 &&
EVP_DigestVerify(m,x->sig,x->siglen,x->msg,sizeof x->msg)>0;
EVP_MD_CTX_free(m);
if(!ok) return 1;
g_sink += 1; /* sink the (successful) verify result */
return 0;
}
static int run_ed25519(const bench_cfg *cfg) {
ed_ctx x; memset(&x,0,sizeof x); memset(x.msg,0xA5,sizeof x.msg);
/* ---- correctness check (ONCE, outside timing): verify MUST succeed ---- */
if (ed_keygen(&x) != 0) die("Ed25519","keygen failed");
if (ed_sign(&x) != 0) die("Ed25519","sign failed");
if (ed_verify(&x) != 0) die("Ed25519","verify failed on a valid signature (broken build)");
measure_out kg={0}, sg={0}, vf={0};
must_measure("Ed25519","keygen",ed_keygen,&x,cfg,&kg);
if (ed_keygen(&x)!=0 || ed_sign(&x)!=0) die("Ed25519","re-priming key+sig failed");
must_measure("Ed25519","sign", ed_sign, &x,cfg,&sg);
must_measure("Ed25519","verify",ed_verify,&x,cfg,&vf);
printf("{\"alg\":\"Ed25519\",\"kind\":\"sig\",\"backend\":\"openssl\",\"classical\":true,\"enabled\":true,");
printf("\"claimed_nist_level\":1,");
printf("\"sizes\":{\"public_key\":32,\"secret_key\":32,\"signature\":64},");
printf("\"operations\":{");
stats_t s;
s=compute_stats(kg.all,kg.all_n); print_stats_json(stdout,"keygen",s,&kg); printf(",");
s=compute_stats(sg.all,sg.all_n); print_stats_json(stdout,"sign",s,&sg); printf(",");
s=compute_stats(vf.all,vf.all_n); print_stats_json(stdout,"verify",s,&vf);
printf("}}\n");
free(kg.all); free(sg.all); free(vf.all);
if(x.key)EVP_PKEY_free(x.key);
return 0;
}
/* ---- main --------------------------------------------------------------- */
static void usage(void){
fprintf(stderr,
"usage: bench_pq --kind kem|sig --alg NAME [options]\n"
" auto-calibration (default): each op is timed to ~--target-time-ms of\n"
" aggregate work, clamped to [--min-samples, --max-iters].\n"
" --target-time-ms N per-op timed-work target (default 250)\n"
" --min-samples N floor on timed iters per rep (default 30)\n"
" --max-iters N ceiling on timed iters per rep (default 20000)\n"
" --reps N independent repetitions (default 5)\n"
" --iters N FIXED-count fallback: exactly N timed iters/rep\n"
" (disables calibration; pairs with --warmup)\n"
" --warmup N warmup iters/rep in fixed-count mode (default 1000)\n");
}
int main(int argc, char **argv) {
const char *kind=NULL, *alg=NULL;
/* fixed-count fallback knobs */
uint64_t warmup=1000, iters=0, reps=5; /* iters=0 => auto-calibrate */
/* auto-calibration knobs */
uint64_t target_time_ms=250, min_samples=30, max_iters=20000;
for (int i=1;i<argc;i++){
if(!strcmp(argv[i],"--kind")&&i+1<argc) kind=argv[++i];
else if(!strcmp(argv[i],"--alg")&&i+1<argc) alg=argv[++i];
else if(!strcmp(argv[i],"--warmup")&&i+1<argc) warmup=strtoull(argv[++i],0,10);
else if(!strcmp(argv[i],"--iters")&&i+1<argc) iters=strtoull(argv[++i],0,10);
else if(!strcmp(argv[i],"--reps")&&i+1<argc) reps=strtoull(argv[++i],0,10);
else if(!strcmp(argv[i],"--target-time-ms")&&i+1<argc) target_time_ms=strtoull(argv[++i],0,10);
else if(!strcmp(argv[i],"--min-samples")&&i+1<argc) min_samples=strtoull(argv[++i],0,10);
else if(!strcmp(argv[i],"--max-iters")&&i+1<argc) max_iters=strtoull(argv[++i],0,10);
else { usage(); return 2; }
}
if(!kind||!alg){ usage(); return 2; }
if(reps < 1) reps = 1;
if(min_samples < 1) min_samples = 1;
if(max_iters < min_samples) max_iters = min_samples;
bench_cfg cfg = {
.fixed_iters = iters, /* >0 => fixed-count mode */
.target_ns = target_time_ms * 1000000ull,
.min_samples = min_samples,
.max_iters = max_iters,
.warmup = warmup,
.reps = reps,
};
if (iters>0)
fprintf(stderr,"[bench_pq] mode=fixed-count reps=%llu warmup=%llu iters=%llu\n",
(unsigned long long)reps,(unsigned long long)warmup,(unsigned long long)iters);
else
fprintf(stderr,"[bench_pq] mode=auto-calibrate reps=%llu target=%llums min_samples=%llu max_iters=%llu\n",
(unsigned long long)reps,(unsigned long long)target_time_ms,
(unsigned long long)min_samples,(unsigned long long)max_iters);
/* PMU cycle availability probe (reported once via stderr-free channel:
* embedded into the JSON header line below). */
const char *pmu_reason=NULL;
int pmu_ok = probe_pmu(&pmu_reason);
fprintf(stderr,"[bench_pq] cycles_available=%d (%s)\n", pmu_ok, pmu_reason);
OQS_init();
int rc;
if(!strcmp(kind,"kem")){
if(!strcmp(alg,"X25519")) rc=run_x25519(&cfg);
else rc=run_kem(alg,&cfg);
} else if(!strcmp(kind,"sig")){
if(!strcmp(alg,"Ed25519")) rc=run_ed25519(&cfg);
else rc=run_sig(alg,&cfg);
} else { usage(); rc=2; }
OQS_destroy();
return rc;
}

245
pq-bench-rpi5/bench/lib/assemble.py Normal file
View File

@ -0,0 +1,245 @@
#!/usr/bin/env python3
"""assemble.py — merge harness outputs + thermal trace + environment metadata
into one results JSON with full provenance.
Inputs (all paths):
--meta meta.env KEY=VALUE run/host facts collected by run.sh
--lock versions.lock toolchain provenance from setup.sh
--features cpu_features.json CPU/crypto-extension detection (from run.sh)
--kemsig kemsig.jsonl one JSON object per algorithm from bench_pq
--tls tls.json output of the TLS harness (optional)
--thermal thermal.csv epoch_s,arm_clock_hz,temp_c,throttled_hex samples
--out results/<host>-<ts>.json
The single most important output field is `is_baseline_grade`: true ONLY on a
real RPi5 with performance governor, core pinning, A76-targeted flags, and no
thermal throttling. Everything else (notably macOS smoke runs) is false, with
reasons recorded — so smoke output can never be mistaken for the baseline.
"""
from __future__ import annotations
import argparse
import json
import os
import statistics
import sys
SCHEMA_VERSION = "1.0.0"
def parse_envfile(path: str) -> dict:
"""Parse KEY=VALUE / KEY="value" lines (shared format of meta.env + versions.lock)."""
out = {}
if not path or not os.path.exists(path):
return out
with open(path) as f:
for line in f:
line = line.strip()
if not line or line.startswith("#") or "=" not in line:
continue
k, v = line.split("=", 1)
v = v.strip()
if len(v) >= 2 and v[0] == v[-1] and v[0] in "\"'":
v = v[1:-1]
out[k.strip()] = v
return out
def load_jsonl(path: str) -> list:
items = []
if not path or not os.path.exists(path):
return items
with open(path) as f:
for line in f:
line = line.strip()
if line:
try:
items.append(json.loads(line))
except json.JSONDecodeError as e:
print(f"[assemble] skipping bad JSONL line: {e}", file=sys.stderr)
return items
def load_json(path: str):
if path and os.path.exists(path):
with open(path) as f:
return json.load(f)
return None
def parse_thermal(path: str) -> dict:
"""Reduce the raw CSV trace to a compact embedded record + summary."""
cols = ["epoch_s", "arm_clock_hz", "temp_c", "throttled_hex"]
samples, temps, clocks = [], [], []
throttling_detected = False
if path and os.path.exists(path):
with open(path) as f:
for line in f:
parts = line.strip().split(",")
if len(parts) != 4:
continue
ep, clk, temp, thr = parts
samples.append([
int(ep) if ep else None,
int(clk) if clk else None,
float(temp) if temp else None,
thr or None,
])
if temp:
temps.append(float(temp))
if clk:
clocks.append(int(clk))
if thr:
try:
v = int(thr, 16)
# bit2 = throttling now, bit18 = throttling has occurred
if v & 0x4 or v & 0x40000:
throttling_detected = True
except ValueError:
pass
def summarize(vals):
if not vals:
return None
return {
"min": min(vals), "max": max(vals),
"mean": round(statistics.fmean(vals), 3),
"samples": len(vals),
}
clock_summary = summarize(clocks)
# Detect frequency droop as a secondary throttling signal.
if clock_summary and clocks:
spread = (max(clocks) - min(clocks)) / max(clocks)
clock_summary["spread_frac"] = round(spread, 4)
return {
"columns": cols,
"samples": samples,
"temp_c": summarize(temps),
"arm_clock_hz": clock_summary,
"throttling_detected": throttling_detected,
}
def to_int(s, default=None):
try:
return int(s)
except (TypeError, ValueError):
return default
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--meta", required=True)
ap.add_argument("--lock", default="")
ap.add_argument("--features", default="")
ap.add_argument("--kemsig", default="")
ap.add_argument("--tls", default="")
ap.add_argument("--thermal", default="")
ap.add_argument("--config", default="")
ap.add_argument("--out", required=True)
args = ap.parse_args()
meta = parse_envfile(args.meta)
lock = parse_envfile(args.lock)
features = load_json(args.features) or {}
kemsig = load_jsonl(args.kemsig)
tls = load_json(args.tls)
thermal = parse_thermal(args.thermal)
is_rpi = meta.get("IS_RPI") == "1"
governor = meta.get("GOVERNOR_AFTER") or meta.get("GOVERNOR_BEFORE") or "unknown"
pinned = meta.get("PINNED") == "1"
cflags_target = lock.get("CFLAGS_TARGET", "unknown")
# ---- the anti-confusion gate -----------------------------------------
baseline_reasons = []
if not is_rpi:
baseline_reasons.append(
f"host is not a Raspberry Pi (model='{meta.get('RPI_MODEL','')}', os={meta.get('OS')})")
if governor != "performance":
baseline_reasons.append(f"CPU governor is '{governor}', not 'performance'")
if not pinned:
baseline_reasons.append("benchmark was not pinned to a dedicated core (no taskset)")
if cflags_target != "cortex-a76":
baseline_reasons.append(f"build flags targeted '{cflags_target}', not cortex-a76")
if thermal.get("throttling_detected"):
baseline_reasons.append("thermal throttling was detected during the run")
is_baseline_grade = len(baseline_reasons) == 0
warnings = []
raw_warn = meta.get("WARNINGS", "")
if raw_warn:
warnings.extend([w for w in raw_warn.split("||") if w])
if not is_baseline_grade:
warnings.append("NOT RPi5-baseline-grade: " + "; ".join(baseline_reasons))
result = {
"schema_version": SCHEMA_VERSION,
"tool_version": meta.get("TOOL_VERSION", "0.1.0"),
"generated_utc": meta.get("TS_END_UTC", ""),
"is_baseline_grade": is_baseline_grade,
"baseline_grade_reasons": baseline_reasons,
"host": {
"hostname": meta.get("HOSTNAME", ""),
"os": meta.get("OS", ""),
"os_pretty": meta.get("OS_PRETTY", ""),
"arch": meta.get("ARCH", ""),
"kernel": meta.get("KERNEL", ""),
"is_rpi": is_rpi,
"rpi_model": meta.get("RPI_MODEL", ""),
"cpu_brand": meta.get("CPU_BRAND", ""),
"ncpu": to_int(meta.get("NCPU")),
"ram_bytes": to_int(meta.get("RAM_BYTES")),
},
"cpu_features": features,
"run": {
"timestamp_start_utc": meta.get("TS_START_UTC", ""),
"timestamp_end_utc": meta.get("TS_END_UTC", ""),
"duration_s": to_int(meta.get("DURATION_S")),
"governor_requested": meta.get("GOVERNOR_REQUESTED", ""),
"governor_before": meta.get("GOVERNOR_BEFORE", ""),
"governor_after": meta.get("GOVERNOR_AFTER", ""),
"bench_core": to_int(meta.get("BENCH_CORE")),
"pinned": pinned,
"taskset_cmd": meta.get("TASKSET_CMD", ""),
# Per-op sizing. In auto-calibration mode warmup_iters/timed_iters are
# chosen per operation (see each entry under kem/sig "operations");
# the run-level target/min/max below describe how they were derived.
"calibration_mode": meta.get("CALIB_MODE", "auto"),
"target_time_ms": to_int(meta.get("TARGET_TIME_MS")),
"min_samples": to_int(meta.get("MIN_SAMPLES")),
"max_iters": to_int(meta.get("MAX_ITERS")),
"warmup_iters": to_int(meta.get("WARMUP")),
"timed_iters": to_int(meta.get("ITERS")),
"repetitions": to_int(meta.get("REPS")),
"cycles_mode": meta.get("CYCLES_MODE", ""),
"cycles_available": meta.get("CYCLES_AVAILABLE") == "1",
"cycles_reason": meta.get("CYCLES_REASON", ""),
},
"toolchain": {
"cc_version": lock.get("CC_VERSION", ""),
"bench_cflags": lock.get("BENCH_CFLAGS", ""),
"cflags_target": cflags_target,
"liboqs_ref": lock.get("LIBOQS_REF", ""),
"liboqs_commit": lock.get("LIBOQS_COMMIT", ""),
"liboqs_opt_defines": lock.get("LIBOQS_OPT_DEFINES", ""),
"openssl": lock.get("OPENSSL_COMMIT", ""),
"oqsprovider_ref": lock.get("OQSPROVIDER_REF", ""),
"oqsprovider_commit": lock.get("OQSPROVIDER_COMMIT", ""),
},
"thermal_trace": thermal,
"warnings": warnings,
"kem": [r for r in kemsig if r.get("kind") == "kem"],
"sig": [r for r in kemsig if r.get("kind") == "sig"],
"tls": tls,
}
os.makedirs(os.path.dirname(args.out) or ".", exist_ok=True)
with open(args.out, "w") as f:
json.dump(result, f, indent=2)
print(args.out)
if __name__ == "__main__":
main()

59
pq-bench-rpi5/bench/lib/list_algs.py Normal file
View File

@ -0,0 +1,59 @@
#!/usr/bin/env python3
"""list_algs.py — expand config.yaml into shell-consumable lines for run.sh.
Modes:
measurement -> KEY=VALUE lines (auto-calib: target/min_samples/max_iters/
reps/cycles_mode; plus warmup/iters fixed-count fallback)
kemsig -> one "kind<TAB>alg<TAB>is_classical" line per algorithm,
baselines first (so charts always have the reference point)
tls -> emits the TLS matrix as JSON on one line
Uses the dependency-free miniyaml parser so no PyYAML is required.
"""
import os
import sys
import json
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import miniyaml # noqa: E402
def main():
mode = sys.argv[1] if len(sys.argv) > 1 else "kemsig"
cfg_path = sys.argv[2] if len(sys.argv) > 2 else "config.yaml"
cfg = miniyaml.load_file(cfg_path)
if mode == "measurement":
m = cfg.get("measurement", {}) or {}
# auto-calibration knobs (default path)
print(f"TARGET_TIME_MS={m.get('target_time_ms', 250)}")
print(f"MIN_SAMPLES={m.get('min_samples', 30)}")
print(f"MAX_ITERS={m.get('max_iters', 20000)}")
print(f"REPS={m.get('repetitions', 5)}")
print(f"CYCLES_MODE={m.get('cycles_mode', 'auto')}")
# fixed-count fallback knobs (used only with ./run.sh --iters)
print(f"WARMUP={m.get('warmup_iters', 1000)}")
print(f"ITERS={m.get('timed_iters', 10000)}")
elif mode == "kemsig":
for kind in ("kem", "sig"):
section = cfg.get(kind, {}) or {}
for grp in ("baseline", "candidates"):
for item in (section.get(grp) or []):
if isinstance(item, dict):
name = item.get("name")
classical = "1" if item.get("classical") else "0"
else:
name, classical = item, "0"
if name:
print(f"{kind}\t{name}\t{classical}")
elif mode == "tls":
print(json.dumps(cfg.get("tls", {})))
else:
sys.exit(f"unknown mode: {mode}")
if __name__ == "__main__":
main()

151
pq-bench-rpi5/bench/lib/miniyaml.py Normal file
View File

@ -0,0 +1,151 @@
"""miniyaml — a dependency-free parser for the restricted YAML subset used by
config.yaml.
We deliberately avoid a PyYAML runtime dependency so the core pipeline runs on a
stock Python 3 (the RPi5 / Mac smoke box both have only stdlib by default).
Supported subset (sufficient for config.yaml):
- nested mappings via indentation (2 spaces per level by convention)
- lists of scalars: "- value"
- lists of mappings: "- key: value" then indented "key: value"
- scalars: int, float, bool (true/false), null/~, quoted or bare strings
- "# comment" to end of line (outside quotes)
It is NOT a general YAML implementation. If a user needs full YAML they can
install PyYAML and set PQB_USE_PYYAML=1.
"""
from __future__ import annotations
import os
import re
def _scalar(tok: str):
s = tok.strip()
if s == "" or s in ("~", "null", "Null", "NULL"):
return None
if (s[0] == '"' and s[-1] == '"') or (s[0] == "'" and s[-1] == "'"):
return s[1:-1]
low = s.lower()
if low in ("true", "yes"):
return True
if low in ("false", "no"):
return False
if re.fullmatch(r"[+-]?\d+", s):
return int(s)
if re.fullmatch(r"[+-]?\d*\.\d+([eE][+-]?\d+)?", s):
return float(s)
return s
def _strip_comment(line: str) -> str:
out, q = [], None
for ch in line:
if q:
out.append(ch)
if ch == q:
q = None
elif ch in ("'", '"'):
q = ch
out.append(ch)
elif ch == "#":
break
else:
out.append(ch)
return "".join(out).rstrip()
def _indent(line: str) -> int:
return len(line) - len(line.lstrip(" "))
def loads(text: str):
# Tokenize into (indent, content) ignoring blank/comment-only lines.
lines = []
for raw in text.splitlines():
c = _strip_comment(raw)
if c.strip() == "":
continue
lines.append((_indent(c), c.strip(), c))
pos = [0]
def parse_block(min_indent: int):
if pos[0] >= len(lines):
return None
indent = lines[pos[0]][0]
if lines[pos[0]][1].startswith("- "):
return parse_list(indent)
return parse_map(indent)
def parse_map(indent: int):
obj = {}
while pos[0] < len(lines):
ind, stripped, _ = lines[pos[0]]
if ind < indent or stripped.startswith("- "):
break
if ind > indent: # malformed; skip
pos[0] += 1
continue
m = re.match(r"^([^:]+):\s*(.*)$", stripped)
if not m:
pos[0] += 1
continue
key, val = m.group(1).strip(), m.group(2)
pos[0] += 1
if val == "":
# nested block or empty
if pos[0] < len(lines) and lines[pos[0]][0] > indent:
obj[key] = parse_block(indent + 1)
else:
obj[key] = None
else:
obj[key] = _scalar(val)
return obj
def parse_list(indent: int):
arr = []
while pos[0] < len(lines):
ind, stripped, _ = lines[pos[0]]
if ind < indent or not stripped.startswith("- "):
break
if ind > indent:
break
item = stripped[2:].strip()
pos[0] += 1
if ":" in item and not (item[0] in "'\""):
# list of mappings — first pair is inline, rest are indented deeper
sub = {}
m = re.match(r"^([^:]+):\s*(.*)$", item)
key, val = m.group(1).strip(), m.group(2)
sub[key] = _scalar(val) if val != "" else None
child_indent = indent + 2
while pos[0] < len(lines) and lines[pos[0]][0] >= child_indent \
and not lines[pos[0]][1].startswith("- "):
ind2, strip2, _ = lines[pos[0]]
m2 = re.match(r"^([^:]+):\s*(.*)$", strip2)
if not m2:
pos[0] += 1
continue
k2, v2 = m2.group(1).strip(), m2.group(2)
pos[0] += 1
sub[k2] = _scalar(v2) if v2 != "" else None
arr.append(sub)
else:
arr.append(_scalar(item))
return arr
return parse_block(0) or {}
def load_file(path: str):
if os.environ.get("PQB_USE_PYYAML") == "1":
import yaml # type: ignore
with open(path) as f:
return yaml.safe_load(f)
with open(path) as f:
return loads(f.read())
if __name__ == "__main__":
import json
import sys
print(json.dumps(load_file(sys.argv[1]), indent=2))

17
pq-bench-rpi5/bench/tls/Makefile Normal file
View File

@ -0,0 +1,17 @@
# Build the TLS handshake harness against the OpenSSL that has oqs-provider.
# make OPENSSL_PREFIX=... (defaults to Homebrew openssl@3, else system)
OPENSSL_PREFIX ?= $(shell brew --prefix openssl@3 2>/dev/null || echo /usr)
CC ?= cc
CFLAGS := -O3 -std=c11 -Wall -Wextra -I$(OPENSSL_PREFIX)/include
LDFLAGS := -L$(OPENSSL_PREFIX)/lib -Wl,-rpath,$(OPENSSL_PREFIX)/lib
LDLIBS := -lssl -lcrypto -lm
bench_tls: bench_tls.c
$(CC) $(CFLAGS) -o $@ $< $(LDFLAGS) $(LDLIBS)
clean:
rm -f bench_tls
.PHONY: clean

194
pq-bench-rpi5/bench/tls/bench_tls.c Normal file
View File

@ -0,0 +1,194 @@
/* ===========================================================================
* bench_tls.c — TLS 1.3 handshake benchmark via the OpenSSL API.
*
* Performs full TLS 1.3 handshakes entirely in-process over a pair of memory
* BIOs (no sockets, no CLI scraping). This gives:
* - clean per-handshake wall-clock latency (no socket/scheduler noise)
* - exact bytes-on-the-wire each direction
* - the precise ClientHello flight size (with a fragmentation note vs MSS)
* - real server-cert signature verification cost (client verifies the chain),
* which is the whole point of sweeping PQ signature algorithms.
*
* bench_tls --group X25519MLKEM768 --ca ca.pem \
* --cert server.pem --key server.key --connections 1000 \
* --label "X25519MLKEM768+mldsa65"
*
* Emits one JSON object. PQ groups/sigs require oqs-provider to be loadable
* (point OPENSSL_MODULES at its directory); if it cannot load, we say so.
* ===========================================================================*/
#define _POSIX_C_SOURCE 200809L
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
#include <math.h>
#include <time.h>
#include <openssl/ssl.h>
#include <openssl/bio.h>
#include <openssl/err.h>
#include <openssl/provider.h>
#define TYPICAL_MSS 1400 /* note fragmentation when ClientHello exceeds this */
static inline uint64_t now_ns(void) {
struct timespec ts; clock_gettime(CLOCK_MONOTONIC, &ts);
return (uint64_t)ts.tv_sec * 1000000000ull + (uint64_t)ts.tv_nsec;
}
static int cmp_u64(const void *a, const void *b) {
uint64_t x = *(const uint64_t *)a, y = *(const uint64_t *)b;
return (x > y) - (x < y);
}
static double pct(const uint64_t *s, uint64_t n, double p) {
if (!n) return 0; if (n == 1) return (double)s[0];
double idx = p * (double)(n - 1); uint64_t lo = (uint64_t)idx; double f = idx - lo;
if (lo + 1 >= n) return (double)s[n - 1];
return (double)s[lo] + f * ((double)s[lo + 1] - (double)s[lo]);
}
/* Shuttle all pending bytes from src's mem BIO into dst's mem BIO.
* Returns bytes moved. */
static size_t pump(BIO *src, BIO *dst) {
char buf[16384]; size_t total = 0; int n;
while ((n = BIO_read(src, buf, sizeof buf)) > 0) {
BIO_write(dst, buf, n);
total += (size_t)n;
}
return total;
}
/* One full handshake. Records bytes each way and ClientHello size (first flight).
* Returns 0 on success. */
static int one_handshake(SSL_CTX *cctx, SSL_CTX *sctx,
size_t *c2s_bytes, size_t *s2c_bytes, size_t *chello) {
SSL *cli = SSL_new(cctx), *srv = SSL_new(sctx);
if (!cli || !srv) { if (cli) SSL_free(cli); if (srv) SSL_free(srv); return -1; }
BIO *cli_rb = BIO_new(BIO_s_mem()), *cli_wb = BIO_new(BIO_s_mem());
BIO *srv_rb = BIO_new(BIO_s_mem()), *srv_wb = BIO_new(BIO_s_mem());
SSL_set_bio(cli, cli_rb, cli_wb); /* takes ownership */
SSL_set_bio(srv, srv_rb, srv_wb);
SSL_set_connect_state(cli);
SSL_set_accept_state(srv);
*c2s_bytes = *s2c_bytes = *chello = 0;
int first_flight = 1, rc = 0;
for (int i = 0; i < 64; i++) {
int r_c = SSL_do_handshake(cli);
size_t moved = pump(cli_wb, srv_rb);
if (first_flight && moved > 0) { *chello = moved; first_flight = 0; }
*c2s_bytes += moved;
(void)r_c;
int r_s = SSL_do_handshake(srv);
*s2c_bytes += pump(srv_wb, cli_rb);
(void)r_s;
if (SSL_is_init_finished(cli) && SSL_is_init_finished(srv)) break;
/* if both stalled with nothing to transfer, it's a failure */
if (BIO_ctrl_pending(cli_wb) == 0 && BIO_ctrl_pending(srv_wb) == 0 &&
!SSL_is_init_finished(cli) && i > 4) { rc = -2; break; }
}
if (!SSL_is_init_finished(cli) || !SSL_is_init_finished(srv)) rc = -2;
SSL_free(cli); SSL_free(srv);
return rc;
}
int main(int argc, char **argv) {
const char *group = "X25519", *ca = NULL, *cert = NULL, *key = NULL, *label = "";
uint64_t connections = 1000, warmup = 20;
for (int i = 1; i < argc; i++) {
if (!strcmp(argv[i], "--group") && i+1<argc) group = argv[++i];
else if (!strcmp(argv[i], "--ca") && i+1<argc) ca = argv[++i];
else if (!strcmp(argv[i], "--cert") && i+1<argc) cert = argv[++i];
else if (!strcmp(argv[i], "--key") && i+1<argc) key = argv[++i];
else if (!strcmp(argv[i], "--connections") && i+1<argc) connections = strtoull(argv[++i],0,10);
else if (!strcmp(argv[i], "--warmup") && i+1<argc) warmup = strtoull(argv[++i],0,10);
else if (!strcmp(argv[i], "--label") && i+1<argc) label = argv[++i];
else { fprintf(stderr,"bad arg %s\n",argv[i]); return 2; }
}
if (!cert || !key) { fprintf(stderr,"--cert and --key required\n"); return 2; }
/* providers: default always; oqs if discoverable (OPENSSL_MODULES) */
OSSL_PROVIDER_load(NULL, "default");
int have_oqs = OSSL_PROVIDER_load(NULL, "oqsprovider") != NULL;
SSL_CTX *cctx = SSL_CTX_new(TLS_client_method());
SSL_CTX *sctx = SSL_CTX_new(TLS_server_method());
if (!cctx || !sctx) { fprintf(stderr,"ctx alloc failed\n"); return 1; }
SSL_CTX_set_min_proto_version(cctx, TLS1_3_VERSION);
SSL_CTX_set_max_proto_version(cctx, TLS1_3_VERSION);
SSL_CTX_set_min_proto_version(sctx, TLS1_3_VERSION);
SSL_CTX_set_max_proto_version(sctx, TLS1_3_VERSION);
int group_ok = SSL_CTX_set1_groups_list(cctx, group) &&
SSL_CTX_set1_groups_list(sctx, group);
int cert_ok = SSL_CTX_use_certificate_chain_file(sctx, cert) == 1 &&
SSL_CTX_use_PrivateKey_file(sctx, key, SSL_FILETYPE_PEM) == 1;
/* client verifies the server chain -> exercises PQ signature verify cost */
int verify_setup = 1;
if (ca) {
if (SSL_CTX_load_verify_locations(cctx, ca, NULL) != 1) verify_setup = 0;
SSL_CTX_set_verify(cctx, SSL_VERIFY_PEER, NULL);
} else {
SSL_CTX_set_verify(cctx, SSL_VERIFY_NONE, NULL);
}
if (!group_ok || !cert_ok || !verify_setup) {
printf("{\"label\":\"%s\",\"group\":\"%s\",\"enabled\":false,\"have_oqs_provider\":%s,"
"\"reason\":\"%s%s%s\"}\n",
label, group, have_oqs?"true":"false",
group_ok?"":"group-not-supported ",
cert_ok?"":"cert-load-failed ",
verify_setup?"":"ca-load-failed");
ERR_print_errors_fp(stderr);
return 0;
}
/* warmup */
size_t c2s, s2c, ch;
for (uint64_t i = 0; i < warmup; i++) {
if (one_handshake(cctx, sctx, &c2s, &s2c, &ch) != 0) {
printf("{\"label\":\"%s\",\"group\":\"%s\",\"enabled\":false,"
"\"have_oqs_provider\":%s,\"reason\":\"handshake failed\"}\n",
label, group, have_oqs?"true":"false");
ERR_print_errors_fp(stderr);
return 0;
}
}
uint64_t *lat = malloc(connections * sizeof(uint64_t));
size_t ch_last = 0, c2s_last = 0, s2c_last = 0;
uint64_t ok = 0;
for (uint64_t i = 0; i < connections; i++) {
uint64_t t0 = now_ns();
int rc = one_handshake(cctx, sctx, &c2s, &s2c, &ch);
uint64_t dt = now_ns() - t0;
if (rc == 0) { lat[ok++] = dt; ch_last = ch; c2s_last = c2s; s2c_last = s2c; }
}
if (ok == 0) { fprintf(stderr,"all handshakes failed\n"); return 1; }
qsort(lat, ok, sizeof(uint64_t), cmp_u64);
double median = pct(lat, ok, 0.5);
double p95 = pct(lat, ok, 0.95);
double mn = (double)lat[0], mx = (double)lat[ok-1];
double sum=0; for (uint64_t i=0;i<ok;i++) sum+=(double)lat[i];
double mean = sum/ok;
double ss=0; for (uint64_t i=0;i<ok;i++){double d=(double)lat[i]-mean; ss+=d*d;}
double stddev = ok>1?sqrt(ss/(ok-1)):0;
double hs_per_sec = median>0 ? 1e9/median : 0;
printf("{\"label\":\"%s\",\"group\":\"%s\",\"enabled\":true,\"have_oqs_provider\":%s,",
label, group, have_oqs?"true":"false");
printf("\"connections\":%llu,\"succeeded\":%llu,", (unsigned long long)connections,(unsigned long long)ok);
printf("\"handshake_latency_ns\":{\"median\":%.1f,\"p95\":%.1f,\"min\":%.1f,\"max\":%.1f,\"mean\":%.1f,\"stddev\":%.1f},",
median,p95,mn,mx,mean,stddev);
printf("\"handshakes_per_sec\":%.1f,", hs_per_sec);
printf("\"bytes_on_wire\":{\"client_to_server\":%zu,\"server_to_client\":%zu,\"total\":%zu},",
c2s_last, s2c_last, c2s_last+s2c_last);
printf("\"client_hello_bytes\":%zu,\"client_hello_fragmented\":%s,\"mss_assumed\":%d}\n",
ch_last, ch_last>TYPICAL_MSS?"true":"false", TYPICAL_MSS);
free(lat);
return 0;
}

157
pq-bench-rpi5/bench/tls/run_tls.sh Executable file
View File

@ -0,0 +1,157 @@
#!/usr/bin/env bash
# =============================================================================
# run_tls.sh — generate PKI and run the TLS 1.3 (KEM-group x signature) matrix.
#
# Always benchmarks the classical Logos baseline (X25519 key exchange + Ed25519
# server auth) using stock OpenSSL. PQ rows additionally require oqs-provider to
# be loadable; if it is not present we record those rows as unavailable (with a
# reason) rather than failing — so this still smoke-tests cleanly on a dev box.
#
# ./run_tls.sh --out tls.json --connections 1000
#
# Honors $PQB_TASKSET (a taskset/numactl prefix) to pin bench_tls to a core.
# =============================================================================
set -euo pipefail
HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT="$(cd "$HERE/../.." && pwd)"
# shellcheck source=setup/lib_platform.sh
source "$ROOT/setup/lib_platform.sh"
LOCK="$ROOT/setup/versions.lock"
# shellcheck disable=SC1090
[ -f "$LOCK" ] && source "$LOCK" || true
pqb_detect_platform
OUT=""; CONNS=1000; WARMUP=20
while [ $# -gt 0 ]; do
case "$1" in
--out) OUT="$2"; shift ;;
--connections) CONNS="$2"; shift ;;
--warmup) WARMUP="$2"; shift ;;
*) pqb_err "unknown arg: $1"; exit 2 ;;
esac
shift
done
[ -n "$OUT" ] || { pqb_err "--out required"; exit 2; }
TASKSET="${PQB_TASKSET:-}"
# ---- choose the OpenSSL that has the provider ------------------------------
OSSL="${OPENSSL_BIN:-$(command -v openssl)}"
OSSL_PREFIX="${OPENSSL_PREFIX:-$(brew --prefix openssl@3 2>/dev/null || echo /usr)}"
PROV_MODULE="${OQSPROVIDER_MODULE:-}"
PROV_ARGS=""
HAVE_OQS=0
if [ -n "$PROV_MODULE" ] && [ -f "$PROV_MODULE" ]; then
export OPENSSL_MODULES="$(dirname "$PROV_MODULE")"
if "$OSSL" list -providers -provider oqsprovider -provider default >/dev/null 2>&1; then
PROV_ARGS="-provider oqsprovider -provider default"
HAVE_OQS=1
pqb_log "oqs-provider available: $PROV_MODULE"
fi
fi
[ "$HAVE_OQS" = 0 ] && pqb_warn "oqs-provider not available — PQ TLS rows will be marked unavailable (classical baseline still runs)"
# ---- build the harness -----------------------------------------------------
make -C "$HERE" OPENSSL_PREFIX="$OSSL_PREFIX" >/dev/null
BENCH="$HERE/bench_tls"
[ -n "$TASKSET" ] && export OPENSSL_MODULES="${OPENSSL_MODULES:-}"
# ---- PKI workspace ---------------------------------------------------------
PKI="$HERE/pki"; rm -rf "$PKI"; mkdir -p "$PKI"
# gen_cert <sig_alg> <out_prefix> [provider] -> CA + server cert/key of that alg
gen_cert() {
local alg="$1" pfx="$2" prov="${3:-}"
local ca_key="$PKI/${pfx}_ca.key" ca_crt="$PKI/${pfx}_ca.pem"
local sv_key="$PKI/${pfx}_server.key" sv_csr="$PKI/${pfx}_server.csr" sv_crt="$PKI/${pfx}_server.pem"
# CA
"$OSSL" req -x509 -new -newkey "$alg" -nodes $prov \
-keyout "$ca_key" -out "$ca_crt" -days 3650 \
-subj "/CN=PQB Test CA ($alg)" >/dev/null 2>&1 || return 1
# server key + CSR + cert signed by CA
"$OSSL" genpkey -algorithm "$alg" $prov -out "$sv_key" >/dev/null 2>&1 || return 1
"$OSSL" req -new -key "$sv_key" $prov -out "$sv_csr" -subj "/CN=localhost" >/dev/null 2>&1 || return 1
"$OSSL" x509 -req -in "$sv_csr" -CA "$ca_crt" -CAkey "$ca_key" $prov \
-out "$sv_crt" -days 3650 -CAcreateserial >/dev/null 2>&1 || return 1
return 0
}
# ---- read config -----------------------------------------------------------
TLS_JSON="$(python3 "$ROOT/bench/lib/list_algs.py" tls "$ROOT/config.yaml")"
read_list() { python3 -c "import json,sys; print('\n'.join(json.loads(sys.argv[1]).get(sys.argv[2],[])))" "$TLS_JSON" "$1"; }
BASE_KEM="$(python3 -c "import json,sys;print(json.loads(sys.argv[1])['baseline']['kem_group'])" "$TLS_JSON")"
BASE_SIG="$(python3 -c "import json,sys;print(json.loads(sys.argv[1])['baseline']['sig_alg'])" "$TLS_JSON")"
# ---- generate certs --------------------------------------------------------
declare -a SIG_OK_ALGS=()
# classical baseline cert (always)
if gen_cert "$BASE_SIG" "base_$BASE_SIG"; then
pqb_log "generated baseline cert ($BASE_SIG)"
else
pqb_err "failed to generate classical baseline cert ($BASE_SIG) — TLS layer cannot run"
echo '{"available":false,"reason":"baseline cert generation failed"}' > "$OUT"; exit 0
fi
if [ "$HAVE_OQS" = 1 ]; then
while IFS= read -r s; do
[ -z "$s" ] && continue
if gen_cert "$s" "pq_$s" "$PROV_ARGS"; then
SIG_OK_ALGS+=("$s"); pqb_log "generated PQ cert ($s)"
else
pqb_warn "could not generate cert for sig alg '$s' (skipping)"
fi
done < <(read_list sig_algs)
fi
# cert path helpers
ca_for() { case "$1" in "$BASE_SIG") echo "$PKI/base_${BASE_SIG}_ca.pem";; *) echo "$PKI/pq_${1}_ca.pem";; esac; }
crt_for() { case "$1" in "$BASE_SIG") echo "$PKI/base_${BASE_SIG}_server.pem";; *) echo "$PKI/pq_${1}_server.pem";; esac; }
key_for() { case "$1" in "$BASE_SIG") echo "$PKI/base_${BASE_SIG}_server.key";; *) echo "$PKI/pq_${1}_server.key";; esac; }
# run one matrix cell -> appends JSON object to $ROWS file
run_cell() {
local kem="$1" sig="$2"
local label="${kem}+${sig}"
# shellcheck disable=SC2086
$TASKSET "$BENCH" --group "$kem" --ca "$(ca_for "$sig")" \
--cert "$(crt_for "$sig")" --key "$(key_for "$sig")" \
--connections "$CONNS" --warmup "$WARMUP" --label "$label" 2>>"$PKI/bench_tls.err"
}
ROWS="$PKI/rows.jsonl"; : > "$ROWS"
# baseline row always first
pqb_log "TLS baseline: $BASE_KEM + $BASE_SIG ($CONNS handshakes)"
run_cell "$BASE_KEM" "$BASE_SIG" >> "$ROWS" || pqb_warn "baseline TLS cell failed"
# PQ matrix: (kem_groups x sig_algs) — only cells whose cert exists
if [ "$HAVE_OQS" = 1 ] && [ "${#SIG_OK_ALGS[@]}" -gt 0 ]; then
while IFS= read -r kem; do
[ -z "$kem" ] && continue
for sig in "${SIG_OK_ALGS[@]}"; do
pqb_log "TLS cell: $kem + $sig"
run_cell "$kem" "$sig" >> "$ROWS" || pqb_warn "cell $kem+$sig failed"
done
done < <(read_list kem_groups)
fi
# ---- assemble tls.json -----------------------------------------------------
python3 - "$ROWS" "$OUT" "$HAVE_OQS" "$BASE_KEM" "$BASE_SIG" <<'PY'
import json,sys
rows_path, out_path, have_oqs, base_kem, base_sig = sys.argv[1:6]
rows=[]
with open(rows_path) as f:
for line in f:
line=line.strip()
if line:
try: rows.append(json.loads(line))
except json.JSONDecodeError: pass
out={
"available": True,
"have_oqs_provider": have_oqs=="1",
"baseline": {"kem_group": base_kem, "sig_alg": base_sig, "label": f"{base_kem}+{base_sig}"},
"matrix": rows,
}
json.dump(out, open(out_path,"w"), indent=2)
print(f"wrote {out_path}: {len(rows)} cells (have_oqs={have_oqs})")
PY

163
pq-bench-rpi5/config.yaml Normal file
View File

@ -0,0 +1,163 @@
# =============================================================================
# pq-bench-rpi5 candidate configuration
#
# This file is the single place users edit to extend the benchmark. The harness
# reads it with a dependency-free YAML subset parser (bench/lib/miniyaml.py), so
# keep to the simple shape used below: top-level maps, lists of scalars, and
# lists of "- key: value" maps. Comments (#) and quotes are fine.
#
# Algorithm names must match liboqs / oqs-provider identifiers exactly. To see
# what your built liboqs supports: ./vendor/liboqs/build/tests/test_kem (lists)
# or check the dashboard's "available algorithms" note after a run.
# =============================================================================
# ---- measurement knobs ------------------------------------------------------
# Default sizing is PER-OPERATION AUTO-CALIBRATION: each op (keygen/encaps/sign/
# ...) is timed long enough to accumulate ~target_time_ms of work, clamped to
# [min_samples, max_iters]. This makes an 18 us ML-KEM keygen and a ~750 ms
# SLH-DSA sign both yield a stable median without a hand-tuned per-alg count:
# - fast ops are bounded by max_iters
# - slow ops are bounded by min_samples (so even ~1 op per target still gets
# enough samples for a stable median/MAD)
# To force FIXED counts instead, pass `./run.sh --iters N` — that disables
# calibration and uses the warmup_iters/timed_iters fallback below.
measurement:
target_time_ms: 250 # auto: aggregate timed work to put on each op
min_samples: 30 # auto: floor on timed iters/rep (stable median+MAD)
max_iters: 20000 # auto: ceiling on timed iters/rep (caps fast ops)
repetitions: 5 # independent repetitions (fresh process each time)
cycles_mode: auto # auto | on | off (PMU userspace cycle counting)
# ---- fixed-count fallback (only used when ./run.sh --iters is given) ----
warmup_iters: 1000 # untimed iterations to settle caches/branch predictors
timed_iters: 10000 # timed iterations per repetition
# ---- security_level (NIST category) ----------------------------------------
# Every candidate carries `security_level: N` — the NIST PQC category the
# parameter set targets (1/2/3/5) — used only to group algorithms on the charts.
# The mapping is PER-SCHEME, so different schemes reach the same security target
# with different-looking sets: a level 1 sitting next to a level 2 is the real
# NIST categorization, not a typo. Notably:
# - ML-KEM-512, Falcon-512, X25519 -> Category 1
# - ML-DSA-44 -> Category 2 (FIPS 204 defines NO
# Category-1 ML-DSA; -44 is its
# smallest set and lands at Cat 2)
# - ML-KEM-768 / ML-DSA-65 -> Category 3
# - ML-KEM-1024 / ML-DSA-87 / Falcon-1024 -> Category 5
# (NIST levels: 1 ~ AES-128, 2 ~ SHA-256 collision, 3 ~ AES-192, 5 ~ AES-256.)
# ---- KEMs ------------------------------------------------------------------
# baseline = the classical reference Logos uses TODAY (drawn on every chart).
kem:
baseline:
- name: X25519
classical: true
security_level: 1 # NIST category (see "security_level" note above)
candidates:
- name: ML-KEM-512
security_level: 1
- name: ML-KEM-768
security_level: 3
- name: ML-KEM-1024
security_level: 5
# Hybrids (classical + PQ) — these are oqs-provider TLS group names; for the
# raw KEM bench they are skipped unless liboqs exposes them as a KEM.
- name: X25519MLKEM768
security_level: 3
hybrid: true
- name: SecP256r1MLKEM768
security_level: 3
hybrid: true
# --- Code-based + conservative-LWE backup candidates (added 2026-06) -------
# Verified enabled in the linked liboqs 0.15.0 build via OQS_KEM_alg_is_enabled
# (oqsconfig.h + runtime cross-check). Keygen/encaps/decaps are reported as
# separate ops so the docs' "keygen is a one-time setup cost" argument can show
# the (slow) McEliece keygen explicitly next to its tiny ciphertext.
#
# Classic McEliece — code-based. Huge public key, TINY ciphertext, very slow
# keygen on the Pi (auto-calibration clamps keygen to min_samples=30, like
# SLH-DSA sign); encaps/decaps are fast and calibrate normally. Standard sets,
# plus the 460896f fast-keygen variant directly next to standard 460896: same
# params and same tiny ciphertext but a faster keygen, so the pair gives the
# keygen trade-off directly (the most useful McEliece data point for the
# migration doc). Other f-variants omitted as redundant.
- name: Classic-McEliece-348864
security_level: 1
- name: Classic-McEliece-460896
security_level: 3
- name: Classic-McEliece-460896f
security_level: 3
- name: Classic-McEliece-6688128
security_level: 5
- name: Classic-McEliece-6960119
security_level: 5
- name: Classic-McEliece-8192128
security_level: 5
# FrodoKEM — conservative (unstructured) LWE. ~100x slower encaps/decaps than
# ML-KEM and ~15x larger ciphertext. AES variant (uses ARM AES instructions,
# which this build compiles in); SHAKE variants also available if needed.
- name: FrodoKEM-640-AES
security_level: 1
- name: FrodoKEM-976-AES
security_level: 3
- name: FrodoKEM-1344-AES
security_level: 5
# HQC-128/192/256: NOT enabled in this liboqs 0.15.0 build (disabled upstream
# after the IND-CCA2 implementation issue; oqsconfig.h has it #undef and the
# runtime OQS_KEM_alg_is_enabled returns 0). Intentionally omitted rather than
# listed-and-disabled. Re-add once linked against a liboqs with HQC re-enabled.
# ---- Signatures ------------------------------------------------------------
sig:
baseline:
- name: Ed25519
classical: true
security_level: 1
candidates:
- name: ML-DSA-44
security_level: 2 # Cat 2, not 1 — FIPS 204 has no Cat-1 ML-DSA (see note above)
- name: ML-DSA-65
security_level: 3
- name: ML-DSA-87
security_level: 5
- name: Falcon-512
security_level: 1
- name: Falcon-1024
security_level: 5
# SLH-DSA (SPHINCS+) — many parameter sets; a representative spread.
- name: SPHINCS+-SHA2-128f-simple
security_level: 1
- name: SPHINCS+-SHA2-128s-simple
security_level: 1
- name: SPHINCS+-SHA2-192f-simple
security_level: 3
- name: SPHINCS+-SHA2-256f-simple
security_level: 5
# --- extend here ---
# ---- TLS 1.3 handshake matrix ----------------------------------------------
# The bench runs (kem_group x sig_alg). The classical baseline pair is ALWAYS
# included as the reference point regardless of what is listed here.
tls:
baseline:
kem_group: X25519
sig_alg: ed25519 # OpenSSL classical auth Logos uses today
# PQ key-exchange groups to test (oqs-provider TLS group names)
kem_groups:
- X25519MLKEM768
- SecP256r1MLKEM768
- mlkem512
- mlkem768
- mlkem1024
# PQ signature algorithms for the server/CA cert (oqs-provider names)
sig_algs:
- mldsa44
- mldsa65
- mldsa87
- falcon512
- sphincssha2128fsimple
connections: 1000 # handshakes timed against a persistent s_server
# ---- future phase (NOT implemented now — hooks only) -----------------------
# zk:
# snark: [] # e.g. groth16, plonk, halo2
# stark: [] # e.g. risc0, winterfell

304
pq-bench-rpi5/dashboard/app.js Normal file
View File

@ -0,0 +1,304 @@
/* pq-bench-rpi5 dashboard — pure client-side, reads a merged.json produced by
* analyze/merge.py. No backend. Renders KEM/sig/TLS charts with the classical
* Logos baseline as a reference line on each. Defaults to baseline-grade (RPi5)
* runs only; a toggle includes macOS/dev smoke runs. */
const LEVEL_COLORS = { 1:"#3bd67a", 2:"#46c0c0", 3:"#3a7bff", 5:"#b06bff", 0:"#888" };
const BASE_COLOR = "#e0533d", PQ_COLOR = "#3a7bff";
let MERGED = null, CHARTS = [];
const $ = (id) => document.getElementById(id);
const nsToMs = (ns) => (ns || 0) / 1e6;
/* compact ms label: more decimals for small values, fewer for large. Guards its
* input — Chart.js may hand a value-label formatter a parsed {x,y} point or a
* non-number; extract the numeric value without coercing an object (Number() on
* a null-prototype object would itself throw), and return "" for anything not
* finite so the value-label draw never throws and halts later charts. */
const fmtMs = (v) => {
const ms = typeof v === "number" ? v
: (v && typeof v === "object") ? (typeof v.y === "number" ? v.y : NaN)
: Number(v);
if (!Number.isFinite(ms)) return "";
return ms>=100?ms.toFixed(0):ms>=10?ms.toFixed(1):ms>=1?ms.toFixed(2):ms.toFixed(3);
};
/* Inline Chart.js plugin: draw each bar's value just above the bar. Enabled
* per-chart via options.plugins.valueLabels.formatter; charts that don't set it
* are untouched (TLS/scatter stay clean). No external dependency, so it works
* even when only the Chart.js core CDN is reachable. */
const valueLabels = {
id: "valueLabels",
afterDatasetsDraw(chart) {
const opt = (chart.options.plugins||{}).valueLabels;
if (!opt || !opt.formatter) return;
const ctx = chart.ctx;
ctx.save();
ctx.fillStyle = "#e6e8ee"; ctx.font = "10px sans-serif";
ctx.textAlign = "center"; ctx.textBaseline = "bottom";
chart.data.datasets.forEach((ds, di) => {
const meta = chart.getDatasetMeta(di);
if (meta.hidden) return;
meta.data.forEach((el, i) => {
const v = ds.data[i];
if (v == null) return;
ctx.fillText(opt.formatter(v, i), el.x, el.y - 3);
});
});
ctx.restore();
}
};
if (window.Chart) Chart.register(valueLabels);
async function boot() {
$("fileInput").addEventListener("change", onFile);
$("includeSmoke").addEventListener("change", render);
$("runSelect").addEventListener("change", render);
try {
const r = await fetch("data/merged.json", { cache: "no-store" });
if (r.ok) { MERGED = await r.json(); afterLoad(); }
else showEmpty("No data/merged.json yet. Run a benchmark, then: " +
"<code>python3 analyze/merge.py results/*.json -o dashboard/data/merged.json</code> " +
"— or load a results file above.");
} catch (e) {
showEmpty("Could not auto-load data/merged.json (open via a local server or use the file picker).");
}
}
function onFile(ev) {
const f = ev.target.files[0]; if (!f) return;
const rd = new FileReader();
rd.onload = () => {
const d = JSON.parse(rd.result);
// accept either a merged file or a single results file
MERGED = d.merged_schema ? d : wrapSingle(d);
afterLoad();
};
rd.readAsText(f);
}
/* wrap a single results JSON into the merged shape so the picker works too */
function wrapSingle(d) {
const rid = `${(d.host||{}).hostname}@${d.generated_utc}`;
const meta = { run_id: rid, hostname:(d.host||{}).hostname, cpu_brand:(d.host||{}).cpu_brand,
is_rpi:(d.host||{}).is_rpi, is_baseline_grade:d.is_baseline_grade };
const kem=[], sig=[], tls=[];
(d.kem||[]).forEach(k=>{ if(k.enabled) Object.entries(k.operations||{}).forEach(([op,st])=>
kem.push({...meta, alg:k.alg, classical:!!k.classical, nist_level:k.claimed_nist_level,
operation:op, median_ns:st.median, sizes:k.sizes})); });
(d.sig||[]).forEach(s=>{ if(s.enabled) Object.entries(s.operations||{}).forEach(([op,st])=>
sig.push({...meta, alg:s.alg, classical:!!s.classical, nist_level:s.claimed_nist_level,
operation:op, median_ns:st.median, sizes:s.sizes})); });
((d.tls||{}).matrix||[]).forEach(c=>{ if(c.enabled) tls.push({...meta, label:c.label, group:c.group,
is_baseline_pair:c.label===((d.tls||{}).baseline||{}).label,
handshakes_per_sec:c.handshakes_per_sec, median_ns:(c.handshake_latency_ns||{}).median,
bytes_total:(c.bytes_on_wire||{}).total, client_hello_bytes:c.client_hello_bytes,
client_hello_fragmented:c.client_hello_fragmented}); });
return { merged_schema:"single", n_runs:1,
runs:[{run_id:rid, host:d.host, is_baseline_grade:d.is_baseline_grade,
baseline_grade_reasons:d.baseline_grade_reasons||[], toolchain:d.toolchain,
cpu_features:d.cpu_features, run:d.run,
thermal_summary:{temp_c:(d.thermal_trace||{}).temp_c,
throttling_detected:(d.thermal_trace||{}).throttling_detected},
generated_utc:d.generated_utc}],
kem, sig, tls };
}
function afterLoad() {
const sel = $("runSelect");
sel.innerHTML = "";
MERGED.runs.forEach(r => {
const o = document.createElement("option");
o.value = r.run_id;
o.textContent = `${(r.host||{}).cpu_brand||"?"} — ${r.is_baseline_grade?"✅ baseline":"⚠ smoke"} — ${r.generated_utc||""}`;
sel.appendChild(o);
});
render();
}
function currentRun() {
const id = $("runSelect").value;
return MERGED.runs.find(r => r.run_id === id) || MERGED.runs[0];
}
function render() {
if (!MERGED) return;
CHARTS.forEach(c => c.destroy()); CHARTS = [];
const run = currentRun();
const includeSmoke = $("includeSmoke").checked;
const allowed = new Set(MERGED.runs
.filter(r => includeSmoke || r.is_baseline_grade)
.map(r => r.run_id));
// chart the selected run if allowed, else fall back to allowed set
const rid = allowed.has(run.run_id) ? run.run_id : null;
const filt = (rows) => rows.filter(r => rid ? r.run_id === rid : allowed.has(r.run_id));
renderBanner(run);
renderEnv(run);
const kem = filt(MERGED.kem), sig = filt(MERGED.sig), tls = filt(MERGED.tls);
barByLevel("kem_keygen", kem, "keygen", "KEM keygen — median latency (ms)");
barByLevel("kem_encaps", kem, "encaps", "KEM encaps — median latency (ms)", "derive");
barByLevel("kem_decaps", kem, "decaps", "KEM decaps — median latency (ms)", "derive");
scatter("kem_scatter", kem, "encaps", "public_key", "KEM size vs speed (encaps)", "public key (B)");
barByLevel("sig_sign", sig, "sign", "Signature sign — median latency (ms)", "sign", true);
barByLevel("sig_verify", sig, "verify", "Signature verify — median latency (ms)", "verify", true);
scatter("sig_scatter", sig, "sign", "signature", "Signature size vs speed (sign)", "signature (B)", true);
tlsThroughput("tls_hs", tls);
tlsClientHello("tls_chello", tls);
}
function renderBanner(run) {
const el = $("quality-banner");
if (run.is_baseline_grade) {
el.className = "banner-good";
el.innerHTML = "✅ RPi5 baseline-grade run — performance governor, core-pinned, " +
"cortex-a76 flags, no thermal throttling.";
} else {
el.className = "banner-warn";
const rs = (run.baseline_grade_reasons||[]).map(r=>`<li>${r}</li>`).join("");
el.innerHTML = "⚠ NOT RPi5 baseline-grade — treat as a pipeline smoke test, not measurement data." +
(rs ? `<ul>${rs}</ul>` : "");
}
}
function renderEnv(run) {
const h = run.host||{}, t = run.toolchain||{}, f = run.cpu_features||{}, rn = run.run||{};
const chip = (label, val) => `<span class="chip"><b>${label}</b> ${val}</span>`;
const sha3 = f.sha3 ? "SHA3 hw ✓" : "SHA3 hw ✗ (Keccak on NEON)";
$("env-summary").innerHTML = [
chip("host", `${h.cpu_brand||"?"} (${h.os_pretty||h.os||"?"})`),
chip("governor", rn.governor_after||"?"),
chip("pinned core", rn.pinned ? rn.bench_core : "no"),
chip("flags", `${t.bench_cflags||"?"} → ${t.cflags_target||"?"}`),
chip("liboqs", `${t.liboqs_ref||"?"} ${(t.liboqs_commit||"").slice(0,8)}`),
chip("crypto-ext", `${f.neon?"NEON ":""}${f.sha2?"SHA2 ":""}${sha3}`),
chip("cycles", rn.cycles_available ? "PMU ✓" : "time-based"),
chip("temp", run.thermal_summary && run.thermal_summary.temp_c
? `${run.thermal_summary.temp_c.mean}°C` + (run.thermal_summary.throttling_detected?" ⚠throttled":"")
: "n/a"),
].join("");
}
/* ---- chart builders ----------------------------------------------------- */
function baselineAnnotation(value, label) {
if (value == null) return {};
return { annotations: { base: {
type:"line", yMin:value, yMax:value, borderColor:BASE_COLOR,
borderWidth:2, borderDash:[6,4],
label:{ display:true, content:label, position:"end",
backgroundColor:BASE_COLOR, font:{size:10} } } } };
}
function barByLevel(canvasId, rows, op, title, baselineOp = op, logY = false) {
const data = rows.filter(r => r.operation === op)
.sort((a,b)=>(a.nist_level||0)-(b.nist_level||0) || a.median_ns-b.median_ns);
if (!data.length) return drawEmpty(canvasId, title);
/* Baseline reference line: the classical row for baselineOp (defaults to this
* chart's own op). KEM encaps/decaps have no classical encaps/decaps op, so
* they map to the X25519 key-agreement (derive) timing instead. */
const base = rows.find(r => r.classical && r.operation === baselineOp);
const ctx = $(canvasId).getContext("2d");
CHARTS.push(new Chart(ctx, {
type:"bar",
data:{ labels:data.map(r=>r.alg),
datasets:[{ label:title,
data:data.map(r=>nsToMs(r.median_ns)),
backgroundColor:data.map(r=> r.classical?BASE_COLOR:(LEVEL_COLORS[r.nist_level]||PQ_COLOR)) }] },
options:{ responsive:true, plugins:{
title:{display:true,text:title,color:"#e6e8ee"},
legend:{display:false},
valueLabels:{ formatter:(v)=>fmtMs(v) },
tooltip:{callbacks:{
label:(it)=>`median ${it.raw.toFixed(4)} ms (${Math.round(it.raw*1e6).toLocaleString()} ns)`,
afterLabel:(it)=>{
const r=data[it.dataIndex]; return `NIST L${r.nist_level} · ${r.classical?"classical baseline":"PQ"}`; }}},
annotation: base ? baselineAnnotation(nsToMs(base.median_ns),
baselineOp === op ? `baseline ${base.alg}` : `baseline ${base.alg} ${base.operation}`) : {} },
scales:{ x:{ticks:{color:"#9aa3b2",maxRotation:50,minRotation:40}},
y:{ type: logY?"logarithmic":"linear",
title:{display:true,text:logY?"ms (log)":"ms",color:"#9aa3b2"},ticks:{color:"#9aa3b2"}} } }
}));
}
function scatter(canvasId, rows, op, sizeKey, title, xlabel, logScale = false) {
const data = rows.filter(r => r.operation === op && r.sizes && r.sizes[sizeKey]);
if (!data.length) return drawEmpty(canvasId, title);
const pts = data.map(r => ({ x:r.sizes[sizeKey], y:nsToMs(r.median_ns), alg:r.alg, classical:r.classical }));
const ctx = $(canvasId).getContext("2d");
CHARTS.push(new Chart(ctx, {
type:"scatter",
data:{ datasets:[{ label:title, data:pts, pointRadius:6,
backgroundColor:pts.map(p=>p.classical?BASE_COLOR:PQ_COLOR) }] },
options:{ responsive:true, plugins:{
title:{display:true,text:title,color:"#e6e8ee"}, legend:{display:false},
tooltip:{callbacks:{label:(it)=>`${it.raw.alg}: ${it.raw.x} B, ${it.raw.y.toFixed(3)} ms`}} },
scales:{ x:{ type: logScale?"logarithmic":"linear",
title:{display:true,text:logScale?`${xlabel} (log)`:xlabel,color:"#9aa3b2"},ticks:{color:"#9aa3b2"}},
y:{ type: logScale?"logarithmic":"linear",
title:{display:true,text:logScale?"median latency (ms, log)":"median latency (ms)",color:"#9aa3b2"},ticks:{color:"#9aa3b2"}} } }
}));
}
function tlsThroughput(canvasId, rows) {
if (!rows.length) return drawEmpty(canvasId, "TLS handshakes/sec — run the TLS layer");
const data = rows.slice().sort((a,b)=>(b.handshakes_per_sec||0)-(a.handshakes_per_sec||0));
const base = data.find(r => r.is_baseline_pair);
const ctx = $(canvasId).getContext("2d");
CHARTS.push(new Chart(ctx, {
type:"bar",
data:{ labels:data.map(r=>r.label),
datasets:[{ label:"handshakes/sec",
data:data.map(r=>r.handshakes_per_sec),
backgroundColor:data.map(r=>r.is_baseline_pair?BASE_COLOR:PQ_COLOR) }] },
options:{ indexAxis:"y", responsive:true, plugins:{
title:{display:true,text:"TLS 1.3 handshake throughput (higher = better)",color:"#e6e8ee"},
legend:{display:false},
annotation: base ? { annotations:{ base:{ type:"line",
xMin:base.handshakes_per_sec, xMax:base.handshakes_per_sec,
borderColor:BASE_COLOR, borderWidth:2, borderDash:[6,4],
label:{display:true,content:`baseline ${base.label}`,position:"end",
backgroundColor:BASE_COLOR,font:{size:10}} } } } : {} },
scales:{ x:{title:{display:true,text:"handshakes/sec",color:"#9aa3b2"},ticks:{color:"#9aa3b2"}},
y:{ticks:{color:"#9aa3b2",font:{size:10}}} } }
}));
}
function tlsClientHello(canvasId, rows) {
if (!rows.length) return drawEmpty(canvasId, "ClientHello size — run the TLS layer");
const data = rows.slice().sort((a,b)=>(b.client_hello_bytes||0)-(a.client_hello_bytes||0));
const base = data.find(r => r.is_baseline_pair);
const ctx = $(canvasId).getContext("2d");
CHARTS.push(new Chart(ctx, {
type:"bar",
data:{ labels:data.map(r=>r.label),
datasets:[{ label:"ClientHello bytes",
data:data.map(r=>r.client_hello_bytes),
backgroundColor:data.map(r=> r.is_baseline_pair?BASE_COLOR
: (r.client_hello_fragmented?"#d98b2b":PQ_COLOR)) }] },
options:{ indexAxis:"y", responsive:true, plugins:{
title:{display:true,text:"ClientHello size (orange = exceeds ~1400B MSS → fragments)",color:"#e6e8ee"},
legend:{display:false},
annotation:{ annotations:{ mss:{ type:"line", xMin:1400, xMax:1400,
borderColor:"#d98b2b", borderWidth:1, borderDash:[4,4],
label:{display:true,content:"~MSS 1400B",position:"start",backgroundColor:"#d98b2b",font:{size:9}} },
...(base?{base:{type:"line",xMin:base.client_hello_bytes,xMax:base.client_hello_bytes,
borderColor:BASE_COLOR,borderWidth:2,borderDash:[6,4],
label:{display:true,content:`baseline ${base.label}`,position:"end",backgroundColor:BASE_COLOR,font:{size:10}}}}:{}) } } },
scales:{ x:{title:{display:true,text:"bytes",color:"#9aa3b2"},ticks:{color:"#9aa3b2"}},
y:{ticks:{color:"#9aa3b2",font:{size:10}}} } }
}));
}
function drawEmpty(canvasId, msg) {
const c = $(canvasId); const ctx = c.getContext("2d");
ctx.clearRect(0,0,c.width,c.height);
ctx.fillStyle = "#9aa3b2"; ctx.font = "13px sans-serif"; ctx.textAlign="center";
ctx.fillText(msg, c.width/2, c.height/2);
}
function showEmpty(html) {
document.querySelector("main").innerHTML = `<div class="empty">${html}</div>`;
}
boot();

4322
pq-bench-rpi5/dashboard/data/merged.json Normal file
View File

File diff suppressed because it is too large Load Diff

68
pq-bench-rpi5/dashboard/index.html Normal file
View File

@ -0,0 +1,68 @@
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>PQ Benchmark — RPi5 baseline</title>
<link rel="stylesheet" href="style.css">
<!-- Chart.js + annotation plugin via CDN (works on GitHub Pages; charts still
render without the annotation line if the CDN is unreachable). -->
<script src="https://cdn.jsdelivr.net/npm/chart.js@4.4.1/dist/chart.umd.min.js"></script>
<script src="https://cdn.jsdelivr.net/npm/chartjs-plugin-annotation@3.0.1/dist/chartjs-plugin-annotation.min.js"></script>
</head>
<body>
<header>
<h1>Post-Quantum Crypto Benchmark <span class="sub">Raspberry Pi 5 baseline</span></h1>
<p class="framing">Migration cost: moving from what Logos uses <b>today</b>
(X25519 + Ed25519) to PQ candidates, on validator-grade hardware. The
classical baseline is drawn as a reference line on every chart.</p>
</header>
<section id="controls">
<label>Dataset:
<input type="file" id="fileInput" accept="application/json">
<span class="hint">or auto-loads <code>data/merged.json</code></span>
</label>
<label class="toggle">
<input type="checkbox" id="includeSmoke"> include non-baseline (smoke / dev) runs
</label>
<label>Run:
<select id="runSelect"></select>
</label>
</section>
<div id="quality-banner"></div>
<div id="env-summary"></div>
<main>
<h2>Key Encapsulation (KEM)</h2>
<div class="chart-grid">
<figure><canvas id="kem_keygen"></canvas></figure>
<figure><canvas id="kem_encaps"></canvas></figure>
<figure><canvas id="kem_decaps"></canvas></figure>
<figure><canvas id="kem_scatter"></canvas></figure>
</div>
<h2>Signatures</h2>
<div class="chart-grid">
<figure><canvas id="sig_sign"></canvas></figure>
<figure><canvas id="sig_verify"></canvas></figure>
<figure><canvas id="sig_scatter"></canvas></figure>
</div>
<h2>TLS 1.3 Handshakes <span class="sub">(KEM group × signature)</span></h2>
<div class="chart-grid">
<figure><canvas id="tls_hs"></canvas></figure>
<figure><canvas id="tls_chello"></canvas></figure>
</div>
</main>
<footer>
<p>Generated by <code>pq-bench-rpi5</code>. Numbers are only RPi5-baseline-grade
when the run banner says so (real Pi 5 · performance governor · core-pinned ·
cortex-a76 flags · no thermal throttling). Everything else is a pipeline smoke test.</p>
</footer>
<script src="app.js"></script>
</body>
</html>

40
pq-bench-rpi5/dashboard/style.css Normal file
View File

@ -0,0 +1,40 @@
:root {
--bg:#0f1115; --panel:#181b22; --ink:#e6e8ee; --muted:#9aa3b2;
--accent:#3bd67a; --base:#e0533d; --pq:#3a7bff; --line:#2a2f3a;
}
* { box-sizing:border-box; }
body { margin:0; font:15px/1.5 -apple-system,Segoe UI,Roboto,sans-serif;
background:var(--bg); color:var(--ink); }
header { padding:24px 28px 8px; }
h1 { margin:0; font-size:22px; }
h1 .sub, h2 .sub { color:var(--muted); font-weight:400; font-size:0.7em; }
.framing { color:var(--muted); max-width:760px; }
h2 { margin:30px 28px 8px; border-bottom:1px solid var(--line); padding-bottom:6px; }
#controls { display:flex; flex-wrap:wrap; gap:18px; align-items:center;
padding:12px 28px; background:var(--panel); margin:8px 0; }
#controls label { color:var(--muted); font-size:14px; }
#controls .hint { font-size:12px; opacity:.7; }
select, input[type=file] { color:var(--ink); background:#0c0e12;
border:1px solid var(--line); border-radius:6px; padding:4px 6px; }
#quality-banner { margin:8px 28px; padding:12px 16px; border-radius:8px; font-weight:600; }
.banner-good { background:#12331f; border:1px solid #1f7a44; color:#7df0a8; }
.banner-warn { background:#3a2410; border:1px solid #aa6a1f; color:#ffce8a; }
.banner-warn ul { font-weight:400; margin:6px 0 0; color:#ffd9a8; }
#env-summary { margin:0 28px 8px; color:var(--muted); font-size:13px;
display:flex; flex-wrap:wrap; gap:6px 18px; }
#env-summary b { color:var(--ink); }
#env-summary .chip { background:var(--panel); padding:3px 9px; border-radius:20px;
border:1px solid var(--line); }
.chart-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(420px,1fr));
gap:18px; padding:14px 28px; }
figure { margin:0; background:var(--panel); border:1px solid var(--line);
border-radius:10px; padding:12px; min-height:320px; }
canvas { max-height:340px; }
footer { color:var(--muted); font-size:12px; padding:18px 28px 40px; max-width:820px; }
code { background:#0c0e12; padding:1px 5px; border-radius:4px; }
.empty { color:var(--muted); padding:30px; text-align:center; }

0
pq-bench-rpi5/results/.gitkeep Normal file
View File

13829
pq-bench-rpi5/results/mehmetmac-20260625T220618Z.json Normal file
View File

File diff suppressed because it is too large Load Diff

10825
pq-bench-rpi5/results/rasberrypi5-20260625T202356Z.json Normal file
View File

File diff suppressed because it is too large Load Diff

290
pq-bench-rpi5/run.sh Executable file
View File

@ -0,0 +1,290 @@
#!/usr/bin/env bash
# =============================================================================
# run.sh — measurement wrapper + orchestrator.
#
# Does the things that make a number credible:
# * sets the CPU governor to `performance` (Linux; warns elsewhere)
# * pins the benchmark to a single isolated core via taskset (core 3 on RPi5;
# core 3 stays clear of CPU0 where the kernel steers IRQs/RPS)
# * logs ARM clock + SoC temperature throughout, embeds the trace in results,
# and warns on thermal throttling
# * runs every candidate from config.yaml, then assembles one results JSON
# stamped with full host + toolchain provenance.
#
# Usage:
# ./run.sh # full run using config.yaml knobs
# ./run.sh --smoke # tiny iteration counts: pipeline smoke test
# ./run.sh --kemsig-only # skip the TLS layer
# ./run.sh --tls-only # only the TLS layer
# ./run.sh --iters N --warmup N --reps N # override measurement knobs
# sudo ./run.sh # needed on Linux to set the governor
# =============================================================================
set -euo pipefail
ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
TOOL_VERSION="0.1.0"
# shellcheck source=setup/lib_platform.sh
source "$ROOT/setup/lib_platform.sh"
# shellcheck source=setup/versions.env
source "$ROOT/setup/versions.env"
LOCK="$ROOT/setup/versions.lock"
[ -f "$LOCK" ] && source "$LOCK" || pqb_warn "no versions.lock — run ./setup/setup.sh first"
pqb_detect_platform
# ---- args ------------------------------------------------------------------
SMOKE=0; DO_KEMSIG=1; DO_TLS=1
OVR_ITERS=""; OVR_WARMUP=""; OVR_REPS=""
while [ $# -gt 0 ]; do
case "$1" in
--smoke) SMOKE=1 ;;
--kemsig-only) DO_TLS=0 ;;
--tls-only) DO_KEMSIG=0 ;;
--no-tls) DO_TLS=0 ;;
--iters) OVR_ITERS="$2"; shift ;;
--warmup) OVR_WARMUP="$2"; shift ;;
--reps) OVR_REPS="$2"; shift ;;
-h|--help) grep '^#' "$0" | sed 's/^# \{0,1\}//'; exit 0 ;;
*) pqb_err "unknown arg: $1"; exit 2 ;;
esac
shift
done
# ---- measurement knobs (config.yaml, overridable) --------------------------
# Sets TARGET_TIME_MS MIN_SAMPLES MAX_ITERS REPS CYCLES_MODE (auto-calibration,
# the default path) plus WARMUP ITERS (the fixed-count fallback).
eval "$(python3 "$ROOT/bench/lib/list_algs.py" measurement "$ROOT/config.yaml")"
# Mode: auto-calibrate per op (default) unless --iters forces a fixed count.
CALIB_MODE="auto"
if [ -n "$OVR_ITERS" ]; then CALIB_MODE="fixed"; ITERS="$OVR_ITERS"; fi
[ -n "$OVR_WARMUP" ] && WARMUP="$OVR_WARMUP"
[ -n "$OVR_REPS" ] && REPS="$OVR_REPS"
if [ "$SMOKE" = 1 ]; then
# Keep auto-calibration (so each op still reaches target) but a single rep, so
# the sweep stays short. This is a pipeline test, NOT measurement data.
REPS=1
pqb_warn "SMOKE MODE: reps=1 — pipeline test only, NOT measurement data"
fi
# Assemble the per-op sizing args passed to every bench_pq invocation.
if [ "$CALIB_MODE" = "fixed" ]; then
BENCH_SIZE_ARGS=(--warmup "$WARMUP" --iters "$ITERS" --reps "$REPS")
pqb_log "sizing: FIXED-count warmup=$WARMUP iters=$ITERS reps=$REPS"
else
BENCH_SIZE_ARGS=(--target-time-ms "$TARGET_TIME_MS" --min-samples "$MIN_SAMPLES" \
--max-iters "$MAX_ITERS" --reps "$REPS")
pqb_log "sizing: AUTO-calibrate target=${TARGET_TIME_MS}ms min_samples=$MIN_SAMPLES max_iters=$MAX_ITERS reps=$REPS"
fi
BENCH_CORE="${BENCH_CORE:-3}"
export PQB_BENCH_CORE="$BENCH_CORE"
# ---- work directory --------------------------------------------------------
HOST="$(pqb_resolve_hostname)"
TS="$(date -u +%Y%m%dT%H%M%SZ)"
WORK="$ROOT/results/.work-$HOST-$TS"
mkdir -p "$WORK" "$ROOT/results"
KEMSIG_OUT="$WORK/kemsig.jsonl"; : > "$KEMSIG_OUT"
TLS_OUT="$WORK/tls.json"
THERMAL="$WORK/thermal.csv"; : > "$THERMAL"
META="$WORK/meta.env"
FEATURES="$WORK/cpu_features.json"
WARN_ACC=""
add_warn() { WARN_ACC="${WARN_ACC:+$WARN_ACC||}$1"; pqb_warn "$1"; }
# ---- governor --------------------------------------------------------------
GOV_BEFORE="$(pqb_get_governor)"
GOV_AFTER="$(pqb_set_governor_performance || true)"
GOV_REQUESTED="performance"
if [ "$GOV_AFTER" != "performance" ]; then
add_warn "governor is '$GOV_AFTER', not 'performance' (need root on Linux, or unsupported on macOS)"
fi
# ---- core pinning ----------------------------------------------------------
TASKSET="$(pqb_taskset_prefix "$BENCH_CORE")"
if [ -n "$TASKSET" ]; then
PINNED=1; pqb_log "pinning to core $BENCH_CORE via: $TASKSET"
else
PINNED=0; add_warn "core pinning unavailable (no taskset/numactl) — results will be noisier"
fi
# ---- thermal sampler (background) ------------------------------------------
SAMPLE_INTERVAL="${SAMPLE_INTERVAL:-1}"
pqb_log "starting thermal/clock sampler (every ${SAMPLE_INTERVAL}s) -> $THERMAL"
( while :; do pqb_sample_thermal >> "$THERMAL" 2>/dev/null; sleep "$SAMPLE_INTERVAL"; done ) &
SAMPLER_PID=$!
disown "$SAMPLER_PID" 2>/dev/null || true # suppress job-control "Terminated" noise on kill
# shellcheck disable=SC2064
trap "kill $SAMPLER_PID 2>/dev/null || true" EXIT
pqb_sample_thermal >> "$THERMAL" 2>/dev/null # one immediate sample
# ---- CPU features ----------------------------------------------------------
pqb_cpu_features_json > "$FEATURES"
pqb_log "cpu features: $(cat "$FEATURES")"
# Report whether Keccak/SHA3 *instruction* acceleration is available + compiled.
SHA3_HW="$(python3 -c "import json;print(json.load(open('$FEATURES'))['sha3'])" 2>/dev/null || echo unknown)"
case "${LIBOQS_OPT_DEFINES:-}" in
*"OQS_USE_ARM_SHA3_INSTRUCTIONS 1"*) SHA3_COMPILED=1 ;;
*) SHA3_COMPILED=0 ;;
esac
pqb_log "Keccak/SHA3: hw_instructions=$SHA3_HW liboqs_compiled_sha3=$SHA3_COMPILED (A76 has no SHA3 ext; Keccak runs on NEON there)"
TS_START="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
START_EPOCH="$(date +%s)"
# ---- build harness if needed -----------------------------------------------
OSSL_PREFIX_FOR_BUILD="${OPENSSL_PREFIX:-$(brew --prefix openssl@3 2>/dev/null || echo /usr)}"
if [ ! -x "$ROOT/bench/kem_sig/bench_pq" ] || [ "$ROOT/bench/kem_sig/bench_pq.c" -nt "$ROOT/bench/kem_sig/bench_pq" ]; then
pqb_log "building bench_pq harness"
make -C "$ROOT/bench/kem_sig" \
LIBOQS_PREFIX="${PREFIX:-$ROOT/vendor/install}" \
OPENSSL_PREFIX="$OSSL_PREFIX_FOR_BUILD" \
BENCH_CFLAGS="${BENCH_CFLAGS:--O3}" >/dev/null
fi
# ---- KEM/sig sweep ---------------------------------------------------------
CYCLES_AVAILABLE=0; CYCLES_REASON="not probed"
if [ "$DO_KEMSIG" = 1 ]; then
if [ "$CALIB_MODE" = "fixed" ]; then
pqb_log "running KEM/sig sweep (fixed: warmup=$WARMUP iters=$ITERS reps=$REPS)"
else
pqb_log "running KEM/sig sweep (auto-calibrate: target=${TARGET_TIME_MS}ms min_samples=$MIN_SAMPLES max_iters=$MAX_ITERS reps=$REPS)"
fi
while IFS=$'\t' read -r kind alg classical; do
[ -z "$alg" ] && continue
pqb_log " $kind $alg"
ERRF="$WORK/err.$kind.$alg.txt"
# shellcheck disable=SC2086
if $TASKSET "$ROOT/bench/kem_sig/bench_pq" --kind "$kind" --alg "$alg" \
"${BENCH_SIZE_ARGS[@]}" >> "$KEMSIG_OUT" 2>"$ERRF"; then
:
else
add_warn "harness failed for $kind $alg (see $ERRF)"
fi
# capture PMU availability from the harness's stderr (first occurrence)
if grep -q 'cycles_available=1' "$ERRF" 2>/dev/null; then
CYCLES_AVAILABLE=1; CYCLES_REASON="$(sed -n 's/.*cycles_available=1 (\(.*\))/\1/p' "$ERRF" | head -1)"
elif [ "$CYCLES_AVAILABLE" = 0 ] && grep -q 'cycles_available=0' "$ERRF" 2>/dev/null; then
CYCLES_REASON="$(sed -n 's/.*cycles_available=0 (\(.*\))/\1/p' "$ERRF" | head -1)"
fi
done < <(python3 "$ROOT/bench/lib/list_algs.py" kemsig "$ROOT/config.yaml")
fi
# ---- TLS layer -------------------------------------------------------------
if [ "$DO_TLS" = 1 ]; then
if [ -x "$ROOT/bench/tls/run_tls.sh" ]; then
TLS_CONNS="$(python3 -c "import json,sys;print(json.loads(sys.argv[1]).get('connections',1000))" \
"$(python3 "$ROOT/bench/lib/list_algs.py" tls "$ROOT/config.yaml")")"
[ "$SMOKE" = 1 ] && TLS_CONNS=50
pqb_log "running TLS handshake matrix ($TLS_CONNS handshakes/cell)"
if PQB_TASKSET="$TASKSET" "$ROOT/bench/tls/run_tls.sh" \
--out "$TLS_OUT" --connections "$TLS_CONNS" >"$WORK/tls.log" 2>&1; then
pqb_log "TLS layer done ($(grep -c '"label"' "$TLS_OUT" 2>/dev/null || echo 0) cells)"
else
add_warn "TLS layer failed or unavailable (see $WORK/tls.log) — continuing without it"
TLS_OUT=""
fi
else
pqb_warn "TLS harness not present yet — skipping (will be added)"
TLS_OUT=""
fi
fi
# ---- stop sampler, gather timing -------------------------------------------
kill "$SAMPLER_PID" 2>/dev/null || true
trap - EXIT
TS_END="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
DURATION=$(( $(date +%s) - START_EPOCH ))
GOV_AFTER_END="$(pqb_get_governor)"
# ---- host facts ------------------------------------------------------------
collect_host_facts() {
local cpu_brand="" ncpu="" ram="" os_pretty="" kernel
kernel="$(uname -r)"
if [ "$PQB_OS" = "macos" ]; then
cpu_brand="$(sysctl -n machdep.cpu.brand_string 2>/dev/null)"
ncpu="$(sysctl -n hw.ncpu 2>/dev/null)"
ram="$(sysctl -n hw.memsize 2>/dev/null)"
os_pretty="macOS $(sw_vers -productVersion 2>/dev/null) ($(sw_vers -buildVersion 2>/dev/null))"
else
# aarch64 /proc/cpuinfo has no 'model name' line, so this grep misses; the
# trailing `|| true` keeps `set -o pipefail` from aborting the run (errexit)
# before the PQB_RPI_MODEL fallback below can supply the brand.
cpu_brand="$(grep -m1 'model name' /proc/cpuinfo 2>/dev/null | sed 's/.*: //' || true)"
[ -z "$cpu_brand" ] && cpu_brand="$PQB_RPI_MODEL"
ncpu="$( (command -v nproc >/dev/null && nproc) || grep -c ^processor /proc/cpuinfo)"
ram="$(( $(grep -m1 MemTotal /proc/meminfo 2>/dev/null | awk '{print $2}') * 1024 ))"
os_pretty="$(. /etc/os-release 2>/dev/null; echo "$PRETTY_NAME")"
fi
{
echo "TOOL_VERSION=$TOOL_VERSION"
echo "HOSTNAME=$HOST"
echo "OS=$PQB_OS"
echo "ARCH=$PQB_ARCH"
echo "KERNEL=$kernel"
echo "OS_PRETTY=\"$os_pretty\""
echo "IS_RPI=$PQB_IS_RPI"
echo "RPI_MODEL=\"$PQB_RPI_MODEL\""
echo "CPU_BRAND=\"$cpu_brand\""
echo "NCPU=$ncpu"
echo "RAM_BYTES=$ram"
echo "GOVERNOR_REQUESTED=$GOV_REQUESTED"
echo "GOVERNOR_BEFORE=$GOV_BEFORE"
echo "GOVERNOR_AFTER=$GOV_AFTER_END"
echo "BENCH_CORE=$BENCH_CORE"
echo "PINNED=$PINNED"
echo "TASKSET_CMD=\"$TASKSET\""
echo "CALIB_MODE=$CALIB_MODE"
echo "TARGET_TIME_MS=$TARGET_TIME_MS"
echo "MIN_SAMPLES=$MIN_SAMPLES"
echo "MAX_ITERS=$MAX_ITERS"
echo "REPS=$REPS"
# warmup/timed_iters are single values only in fixed-count mode; in auto mode
# they are chosen per-op and recorded in each operation's JSON instead.
if [ "$CALIB_MODE" = "fixed" ]; then
echo "WARMUP=$WARMUP"
echo "ITERS=$ITERS"
else
echo "WARMUP="
echo "ITERS="
fi
echo "CYCLES_MODE=$CYCLES_MODE"
echo "CYCLES_AVAILABLE=$CYCLES_AVAILABLE"
echo "CYCLES_REASON=\"$CYCLES_REASON\""
echo "TS_START_UTC=$TS_START"
echo "TS_END_UTC=$TS_END"
echo "DURATION_S=$DURATION"
echo "WARNINGS=\"$WARN_ACC\""
} > "$META"
}
collect_host_facts
# ---- assemble final results JSON -------------------------------------------
OUT="$ROOT/results/${HOST}-${TS}.json"
python3 "$ROOT/bench/lib/assemble.py" \
--meta "$META" --lock "$LOCK" --features "$FEATURES" \
--kemsig "$KEMSIG_OUT" ${TLS_OUT:+--tls "$TLS_OUT"} \
--thermal "$THERMAL" --config "$ROOT/config.yaml" \
--out "$OUT" >/dev/null
# ---- summary ---------------------------------------------------------------
echo
pqb_log "================ RUN COMPLETE ================"
python3 - "$OUT" <<'PY'
import json,sys
d=json.load(open(sys.argv[1]))
g=d["is_baseline_grade"]
print(f" results: {sys.argv[1]}")
print(f" host: {d['host']['cpu_brand']} ({d['host']['os_pretty']})")
print(f" baseline-grade (RPi5): {g}")
if not g:
for r in d['baseline_grade_reasons']:
print(f" - {r}")
tt=d['thermal_trace']
print(f" thermal: {tt.get('temp_c')} throttling={tt.get('throttling_detected')}")
print(f" kem algos: {sum(1 for x in d['kem'] if x.get('enabled'))} enabled / {len(d['kem'])}")
print(f" sig algos: {sum(1 for x in d['sig'] if x.get('enabled'))} enabled / {len(d['sig'])}")
print(f" cycles available: {d['run']['cycles_available']}")
PY
pqb_log "keep raw work dir? -> $WORK (safe to delete)"

227
pq-bench-rpi5/setup/lib_platform.sh Normal file
View File

@ -0,0 +1,227 @@
# shellcheck shell=bash
# =============================================================================
# lib_platform.sh — portable platform abstraction
#
# Sourced by setup/setup.sh and run.sh. Every operation that differs between the
# RPi5 (Debian/Ubuntu aarch64) measurement target and the macOS/Apple-Silicon
# dev box is funneled through one of these functions, so the *identical* codebase
# runs unchanged on both. Where a capability does not exist on a platform
# (governor control, core pinning, on-die thermal sensors), the function degrades
# gracefully and the caller records that it was unavailable — it never silently
# pretends the action happened.
# =============================================================================
# ---- platform detection ----------------------------------------------------
# Sets: PQB_OS (macos|linux), PQB_ARCH, PQB_IS_RPI (1|0), PQB_RPI_MODEL
pqb_detect_platform() {
PQB_ARCH="$(uname -m)"
case "$(uname -s)" in
Darwin) PQB_OS="macos" ;;
Linux) PQB_OS="linux" ;;
*) PQB_OS="unknown" ;;
esac
PQB_IS_RPI=0
PQB_RPI_MODEL=""
if [ "$PQB_OS" = "linux" ] && [ -r /proc/device-tree/model ]; then
# /proc/device-tree/model is NUL-terminated
PQB_RPI_MODEL="$(tr -d '\0' < /proc/device-tree/model 2>/dev/null)"
case "$PQB_RPI_MODEL" in
*"Raspberry Pi"*) PQB_IS_RPI=1 ;;
esac
fi
export PQB_OS PQB_ARCH PQB_IS_RPI PQB_RPI_MODEL
}
# ---- friendly logging ------------------------------------------------------
pqb_log() { printf '\033[1;34m[pqb]\033[0m %s\n' "$*" >&2; }
pqb_warn() { printf '\033[1;33m[pqb WARN]\033[0m %s\n' "$*" >&2; }
pqb_err() { printf '\033[1;31m[pqb ERR]\033[0m %s\n' "$*" >&2; }
# ---- hostname resolution ---------------------------------------------------
# A "good" host id is non-empty, not localhost, and not an avahi/macOS
# auto-assigned "unknown<hexMAC>" placeholder (which is what shows up when no
# real hostname is set — that produced the ugly results filename before).
_pqb_good_host() {
local h="$1"
[ -n "$h" ] || return 1
case "$h" in
localhost|localhost.*) return 1 ;;
esac
printf '%s' "$h" | grep -Eq '^[Uu]nknown[0-9a-fA-F]{6,}$' && return 1
return 0
}
# Resolve a readable, stable host identifier, falling through:
# $HOSTNAME -> `hostname` -> hostnamectl --static (Linux) /
# scutil --get LocalHostName (macOS) -> short machine id (last resort).
# Domain suffixes (.home/.local/...) are stripped. On the RPi5 this yields the
# actual pi hostname; on this Mac it falls through to LocalHostName.
pqb_resolve_hostname() {
local cands=() c h
cands+=("${HOSTNAME:-}")
cands+=("$(hostname 2>/dev/null || true)")
if [ "${PQB_OS:-}" = "linux" ]; then
cands+=("$(hostnamectl --static 2>/dev/null || true)")
elif [ "${PQB_OS:-}" = "macos" ]; then
cands+=("$(scutil --get LocalHostName 2>/dev/null || true)")
fi
for c in "${cands[@]}"; do
h="${c%%.*}" # strip domain suffix
if _pqb_good_host "$h"; then echo "$h"; return 0; fi
done
# last resort: a short, stable machine id so files never collide as "unknown"
local mid=""
if [ -r /etc/machine-id ]; then
mid="$(cut -c1-12 /etc/machine-id 2>/dev/null)"
elif [ "${PQB_OS:-}" = "macos" ]; then
mid="$(ioreg -rd1 -c IOPlatformExpertDevice 2>/dev/null \
| awk -F'"' '/IOPlatformUUID/{print $4}' | tr -d '-' | cut -c1-12)"
fi
echo "host-${mid:-unknown}"
}
# ---- CPU governor ----------------------------------------------------------
# Returns 0 if it set 'performance', 1 if unavailable. Prints the governor it
# left the system in on stdout.
pqb_set_governor_performance() {
if [ "$PQB_OS" = "linux" ] && [ -d /sys/devices/system/cpu/cpu0/cpufreq ]; then
local ok=1 g
for g in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do
[ -w "$g" ] || { ok=0; continue; }
echo performance > "$g" 2>/dev/null || ok=0
done
if [ "$ok" = 1 ]; then
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor 2>/dev/null
return 0
fi
# try cpupower as a fallback (may need sudo)
if command -v cpupower >/dev/null 2>&1 && cpupower frequency-set -g performance >/dev/null 2>&1; then
echo performance; return 0
fi
pqb_warn "could not set governor to performance (need root? try: sudo ./run.sh)"
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor 2>/dev/null || echo "unknown"
return 1
fi
# macOS / other: no userspace governor control.
echo "unavailable"
return 1
}
pqb_get_governor() {
if [ "$PQB_OS" = "linux" ] && [ -r /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor ]; then
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
else
echo "unavailable"
fi
}
# ---- core pinning ----------------------------------------------------------
# pqb_taskset_prefix <core> -> echoes a command prefix to pin to that core, or
# empty string if pinning is unavailable (caller warns).
pqb_taskset_prefix() {
local core="$1"
if command -v taskset >/dev/null 2>&1; then
echo "taskset -c $core"
elif command -v numactl >/dev/null 2>&1; then
echo "numactl --physcpubind=$core"
else
echo "" # no pinning available (e.g. macOS)
fi
}
# ---- thermal / clock sampling ----------------------------------------------
# pqb_sample_thermal -> one CSV line: epoch_s,arm_clock_hz,temp_c,throttled_hex
# Fields that cannot be read on a platform are emitted as empty (no fake zeros).
pqb_sample_thermal() {
local ts clk temp thr
ts="$(date +%s)"
clk=""; temp=""; thr=""
if command -v vcgencmd >/dev/null 2>&1; then
# Raspberry Pi: authoritative SoC sensors.
clk="$(vcgencmd measure_clock arm 2>/dev/null | sed -n 's/.*=//p')"
temp="$(vcgencmd measure_temp 2>/dev/null | sed -n "s/temp=\([0-9.]*\).*/\1/p")"
thr="$(vcgencmd get_throttled 2>/dev/null | sed -n 's/.*=//p')"
elif [ "$PQB_OS" = "linux" ]; then
# Generic Linux fallback (cpufreq + thermal_zone).
local f
f=/sys/devices/system/cpu/cpu${PQB_BENCH_CORE:-0}/cpufreq/scaling_cur_freq
[ -r "$f" ] && clk="$(( $(cat "$f") * 1000 ))" # kHz -> Hz
if [ -r /sys/class/thermal/thermal_zone0/temp ]; then
local milli; milli="$(cat /sys/class/thermal/thermal_zone0/temp)"
temp="$(awk -v m="$milli" 'BEGIN{printf "%.1f", m/1000}')"
fi
fi
# macOS: live per-core freq/temp require sudo powermetrics; we intentionally
# leave them empty rather than emit misleading values. (Smoke test only.)
printf '%s,%s,%s,%s\n' "$ts" "$clk" "$temp" "$thr"
}
# pqb_throttled_active <throttled_hex> -> 0 if thermal throttling currently/has
# occurred, 1 otherwise. RPi get_throttled bit 0 = under-voltage now,
# bit 1 = arm freq capped now, bit 2 = currently throttled,
# bit 3 = soft temp limit active (and bits 16-19 = "has occurred" latches).
pqb_throttled_active() {
local hex="${1#0x}"
[ -z "$hex" ] && return 1
local val=$(( 16#$hex ))
# bit2 (throttling now) or bit18 (throttling has occurred)
if [ $(( val & 0x4 )) -ne 0 ] || [ $(( val & 0x40000 )) -ne 0 ]; then
return 0
fi
return 1
}
# ---- CPU feature / crypto-extension detection ------------------------------
# Echoes a JSON object describing NEON + SHA3/SHA512 acceleration. Consumed
# verbatim by env metadata so results record whether Keccak accel is in use.
pqb_cpu_features_json() {
local neon=false sha2=false sha3=false sha512=false aes=false pmull=false src="unknown"
if [ "$PQB_OS" = "linux" ] && [ -r /proc/cpuinfo ]; then
src="/proc/cpuinfo"
local feats; feats="$(grep -m1 -i '^Features' /proc/cpuinfo | tr 'A-Z' 'a-z')"
case "$feats" in *" asimd"*|*"neon"*) neon=true;; esac
case "$feats" in *" sha2"*) sha2=true;; esac
case "$feats" in *" sha3"*) sha3=true;; esac
case "$feats" in *" sha512"*) sha512=true;; esac
case "$feats" in *" aes"*) aes=true;; esac
case "$feats" in *" pmull"*) pmull=true;; esac
elif [ "$PQB_OS" = "macos" ]; then
src="sysctl"
neon=true # all Apple Silicon has NEON/ASIMD
[ "$(sysctl -n hw.optional.arm.FEAT_SHA256 2>/dev/null)" = 1 ] && sha2=true
[ "$(sysctl -n hw.optional.arm.FEAT_SHA3 2>/dev/null)" = 1 ] && sha3=true
[ "$(sysctl -n hw.optional.arm.FEAT_SHA512 2>/dev/null)" = 1 ] && sha512=true
[ "$(sysctl -n hw.optional.arm.FEAT_AES 2>/dev/null)" = 1 ] && aes=true
[ "$(sysctl -n hw.optional.arm.FEAT_PMULL 2>/dev/null)" = 1 ] && pmull=true
fi
printf '{"source":"%s","neon":%s,"sha2":%s,"sha3":%s,"sha512":%s,"aes":%s,"pmull":%s}' \
"$src" "$neon" "$sha2" "$sha3" "$sha512" "$aes" "$pmull"
}
# ---- package installation --------------------------------------------------
# pqb_install_build_deps -> installs compiler/cmake/openssl headers per platform.
pqb_install_build_deps() {
if [ "$PQB_OS" = "macos" ]; then
command -v brew >/dev/null 2>&1 || { pqb_err "Homebrew required on macOS: https://brew.sh"; return 1; }
pqb_log "installing build deps via Homebrew"
brew install cmake ninja openssl@3 git python3 >/dev/null || true
elif [ "$PQB_OS" = "linux" ]; then
if command -v apt-get >/dev/null 2>&1; then
pqb_log "installing build deps via apt"
local SUDO=""; [ "$(id -u)" -ne 0 ] && SUDO="sudo"
$SUDO apt-get update -qq
# linux-cpupower provides the `cpupower` binary used by
# pqb_set_governor_performance. (Older releases shipped cpufrequtils, which
# was dropped in Debian 13/trixie — cpupower is the supported replacement.)
$SUDO apt-get install -y -qq \
build-essential cmake ninja-build git python3 perl \
libssl-dev pkg-config astyle doxygen \
linux-cpupower util-linux >/dev/null
else
pqb_warn "no apt-get found; install cmake/ninja/gcc/libssl-dev manually"
fi
fi
}

205
pq-bench-rpi5/setup/setup.sh Executable file
View File

@ -0,0 +1,205 @@
#!/usr/bin/env bash
# =============================================================================
# setup.sh — build + pin the full PQ toolchain from scratch.
#
# ./setup/setup.sh # everything: deps, liboqs, openssl(if needed), oqs-provider
# ./setup/setup.sh liboqs # just liboqs
# ./setup/setup.sh openssl # just openssl (forced from source)
# ./setup/setup.sh provider # just oqs-provider
# ./setup/setup.sh deps # just OS packages
#
# Everything is installed under ./vendor/install (no system pollution). The exact
# resolved git commits + the optimization flags actually used are written to
# setup/versions.lock, which run.sh stamps into every results JSON.
#
# Identical flags for every candidate: -O3 -mcpu=cortex-a76 on the RPi5. On a
# non-A76 host (the macOS smoke box) we fall back to -O3 and RECORD that, so
# smoke-test numbers can never masquerade as the RPi5 baseline.
# =============================================================================
set -euo pipefail
HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT="$(cd "$HERE/.." && pwd)"
# shellcheck source=setup/lib_platform.sh
source "$HERE/lib_platform.sh"
# shellcheck source=setup/versions.env
source "$HERE/versions.env"
pqb_detect_platform
VENDOR="$ROOT/vendor"
SRC="$VENDOR/src"
PREFIX="$VENDOR/install"
mkdir -p "$SRC" "$PREFIX"
JOBS="$( (command -v nproc >/dev/null && nproc) || sysctl -n hw.ncpu 2>/dev/null || echo 4)"
# ---- decide the real optimization flags for THIS host ----------------------
# We only use -mcpu=cortex-a76 if the compiler accepts it AND we're on aarch64.
choose_cflags() {
local cc="${CC:-cc}" probe="$SRC/.flagprobe.c"
echo 'int main(void){return 0;}' > "$probe"
if [ "$PQB_ARCH" = "aarch64" ] && $cc $TARGET_CFLAGS_RPI5 "$probe" -o "$probe.out" 2>/dev/null; then
BENCH_CFLAGS="$TARGET_CFLAGS_RPI5"; CFLAGS_TARGET="cortex-a76"
else
BENCH_CFLAGS="$TARGET_CFLAGS_FALLBACK"; CFLAGS_TARGET="generic-fallback"
fi
rm -f "$probe" "$probe.out"
}
cc_version_string() {
local cc="${CC:-cc}"
"$cc" --version 2>/dev/null | head -1
}
git_pin() { # repo ref destdir
local repo="$1" ref="$2" dest="$3"
if [ -d "$dest/.git" ]; then
pqb_log "updating $(basename "$dest") -> $ref"
git -C "$dest" fetch -q --depth 1 origin "$ref" || git -C "$dest" fetch -q --tags origin
else
pqb_log "cloning $(basename "$dest") @ $ref"
git clone -q --depth 1 --branch "$ref" "$repo" "$dest" 2>/dev/null \
|| git clone -q "$repo" "$dest"
fi
git -C "$dest" checkout -q "$ref" 2>/dev/null || true
git -C "$dest" rev-parse HEAD
}
# ---------------------------------------------------------------------------
build_liboqs() {
choose_cflags
local dest="$SRC/liboqs" commit
commit="$(git_pin "$LIBOQS_REPO" "$LIBOQS_REF" "$dest")"
pqb_log "building liboqs ($LIBOQS_REF @ ${commit:0:12}) flags: $BENCH_CFLAGS"
# OQS_DIST_BUILD=OFF -> native build for the fixed target (no runtime CPU
# dispatch), so -mcpu=cortex-a76 fully drives codegen. The AArch64-optimized
# ML-KEM (mlkem-native) and AArch64 asm backends are enabled by default on
# aarch64 when DIST_BUILD is OFF (compile-time CPU features); verified post-build.
local GEN=(); command -v ninja >/dev/null 2>&1 && GEN=(-G Ninja)
cmake -S "$dest" -B "$dest/build" ${GEN[@]+"${GEN[@]}"} \
-DCMAKE_INSTALL_PREFIX="$PREFIX" \
-DCMAKE_BUILD_TYPE=Release \
-DOQS_DIST_BUILD=OFF \
-DOQS_BUILD_ONLY_LIB=OFF \
-DBUILD_SHARED_LIBS=ON \
-DCMAKE_C_FLAGS="$BENCH_CFLAGS" >/dev/null
cmake --build "$dest/build" --parallel "$JOBS" >/dev/null
cmake --install "$dest/build" >/dev/null
# Prove the optimized backend: capture the aarch64/native defines from the
# generated build config so versions.lock can show what was actually compiled.
local cfg="$dest/build/include/oqs/oqsconfig.h"
LIBOQS_OPT_DEFINES="(oqsconfig.h not found)"
if [ -r "$cfg" ]; then
# strip embedded double-quotes so the value stays valid in versions.lock
LIBOQS_OPT_DEFINES="$(grep -Ei 'AARCH64|ARM|_ASM|MLKEM_NATIVE|OPT_TARGET|CPU_EXT' "$cfg" \
| grep -i 'define' | sed 's/^#define //' | tr -d '"' | tr '\n' ';' || true)"
fi
LIBOQS_COMMIT="$commit"
}
# ---------------------------------------------------------------------------
locate_or_build_openssl() {
# Prefer an existing >= 3.5 openssl (Homebrew on macOS, distro on Linux) unless
# BUILD_OPENSSL=1. PQ sig certs for TLS only need >= 3.5.0.
local want_major=3 want_minor=5
if [ "${1:-}" != "force" ] && [ "${BUILD_OPENSSL:-0}" != 1 ]; then
local cand
for cand in "$(command -v openssl || true)" /opt/homebrew/opt/openssl@3/bin/openssl /usr/bin/openssl; do
[ -x "$cand" ] || continue
local v; v="$("$cand" version 2>/dev/null | awk '{print $2}')"
# NB: assign on separate lines. A single `local a=.. b=.. c="${b..}"` makes
# bash 5.2 declare all names (unset) *before* expanding any RHS, so the
# reference to `rest` here trips `set -u` (unbound variable) on the Pi.
local maj rest min
maj="${v%%.*}"; rest="${v#*.}"; min="${rest%%.*}"
if [ "${maj:-0}" -gt "$want_major" ] 2>/dev/null || \
{ [ "${maj:-0}" -eq "$want_major" ] && [ "${min:-0}" -ge "$want_minor" ]; } 2>/dev/null; then
OPENSSL_BIN="$cand"
OPENSSL_PREFIX="$(dirname "$(dirname "$cand")")"
OPENSSL_COMMIT="system:$v"
pqb_log "using existing OpenSSL $v at $cand"
return 0
fi
done
fi
pqb_log "building OpenSSL $OPENSSL_REF from source"
local dest="$SRC/openssl" commit
commit="$(git_pin "$OPENSSL_REPO" "$OPENSSL_REF" "$dest")"
( cd "$dest" && ./Configure --prefix="$PREFIX" --openssldir="$PREFIX/ssl" shared \
&& make -j"$JOBS" >/dev/null && make install_sw >/dev/null )
OPENSSL_BIN="$PREFIX/bin/openssl"
OPENSSL_PREFIX="$PREFIX"
OPENSSL_COMMIT="$commit"
}
# ---------------------------------------------------------------------------
build_oqs_provider() {
[ -n "${OPENSSL_PREFIX:-}" ] || locate_or_build_openssl
local dest="$SRC/oqs-provider" commit
commit="$(git_pin "$OQSPROVIDER_REPO" "$OQSPROVIDER_REF" "$dest")"
pqb_log "building oqs-provider ($OQSPROVIDER_REF @ ${commit:0:12})"
local GEN=(); command -v ninja >/dev/null 2>&1 && GEN=(-G Ninja)
cmake -S "$dest" -B "$dest/build" ${GEN[@]+"${GEN[@]}"} \
-DCMAKE_INSTALL_PREFIX="$PREFIX" \
-DCMAKE_BUILD_TYPE=Release \
-DOPENSSL_ROOT_DIR="$OPENSSL_PREFIX" \
-Dliboqs_DIR="$PREFIX/lib/cmake/liboqs" \
-DCMAKE_C_FLAGS="${BENCH_CFLAGS:-$TARGET_CFLAGS_FALLBACK}" >/dev/null
cmake --build "$dest/build" --parallel "$JOBS" >/dev/null
cmake --install "$dest/build" >/dev/null 2>&1 || true
# Provider .so lands under .../lib/ossl-modules or .../oqsprovider
OQSPROVIDER_MODULE="$(find "$PREFIX" "$dest/build" -name 'oqsprovider.*' \( -name '*.so' -o -name '*.dylib' \) 2>/dev/null | head -1)"
OQSPROVIDER_COMMIT="$commit"
}
# ---------------------------------------------------------------------------
write_lock() {
choose_cflags 2>/dev/null || true
local lock="$HERE/versions.lock"
{
echo "# Auto-generated by setup.sh — exact toolchain provenance. Stamped into results JSON."
echo "PQB_BUILD_HOST_OS=$PQB_OS"
echo "PQB_BUILD_HOST_ARCH=$PQB_ARCH"
echo "PQB_IS_RPI=$PQB_IS_RPI"
echo "PQB_RPI_MODEL=\"${PQB_RPI_MODEL}\""
echo "BENCH_CFLAGS=\"${BENCH_CFLAGS:-unknown}\""
echo "CFLAGS_TARGET=\"${CFLAGS_TARGET:-unknown}\""
echo "CC_VERSION=\"$(cc_version_string)\""
echo "LIBOQS_REF=\"$LIBOQS_REF\""
echo "LIBOQS_COMMIT=\"${LIBOQS_COMMIT:-not-built}\""
echo "LIBOQS_OPT_DEFINES=\"${LIBOQS_OPT_DEFINES:-}\""
echo "OPENSSL_BIN=\"${OPENSSL_BIN:-}\""
echo "OPENSSL_PREFIX=\"${OPENSSL_PREFIX:-}\""
echo "OPENSSL_COMMIT=\"${OPENSSL_COMMIT:-not-built}\""
echo "OQSPROVIDER_REF=\"$OQSPROVIDER_REF\""
echo "OQSPROVIDER_COMMIT=\"${OQSPROVIDER_COMMIT:-not-built}\""
echo "OQSPROVIDER_MODULE=\"${OQSPROVIDER_MODULE:-}\""
echo "PREFIX=\"$PREFIX\""
} > "$lock"
pqb_log "wrote $lock"
cat "$lock" >&2
}
# ---- dispatch --------------------------------------------------------------
main() {
local what="${1:-all}"
case "$what" in
deps) pqb_install_build_deps ;;
liboqs) build_liboqs; write_lock ;;
openssl) locate_or_build_openssl force; write_lock ;;
provider) build_oqs_provider; write_lock ;;
all)
pqb_install_build_deps
build_liboqs
locate_or_build_openssl
build_oqs_provider
write_lock
pqb_log "setup complete. Next: ./run.sh --smoke"
;;
*) pqb_err "unknown target: $what (deps|liboqs|openssl|provider|all)"; exit 2 ;;
esac
}
main "$@"

40
pq-bench-rpi5/setup/versions.env Normal file
View File

@ -0,0 +1,40 @@
# =============================================================================
# Pinned upstream versions for the PQ benchmark toolchain.
#
# These are the *intended* refs. setup/setup.sh clones each at the tag below
# and then records the *actually resolved* commit hash into setup/versions.lock,
# which is stamped verbatim into every results JSON. That way a results file is
# always traceable to exact source, even if a tag is ever re-pointed upstream.
#
# Override any of these from the environment, e.g.:
# LIBOQS_REF=main ./setup/setup.sh
# =============================================================================
# liboqs: KEM/signature implementations. >= 0.15.0 ships the AArch64-optimized
# ML-KEM backend (mlkem-native). We enable it explicitly in setup.sh.
LIBOQS_REPO="${LIBOQS_REPO:-https://github.com/open-quantum-safe/liboqs.git}"
LIBOQS_REF="${LIBOQS_REF:-0.15.0}"
# OpenSSL >= 3.5.0 — required for PQ signatures in TLS 1.3 server auth.
# On macOS dev boxes we prefer the Homebrew openssl@3 if it is already >= 3.5;
# setup.sh only builds OpenSSL from source when the system one is too old or
# BUILD_OPENSSL=1 is set.
OPENSSL_REPO="${OPENSSL_REPO:-https://github.com/openssl/openssl.git}"
OPENSSL_REF="${OPENSSL_REF:-openssl-3.5.0}"
# oqs-provider >= 0.9.0 — wires liboqs algorithms into OpenSSL as a provider,
# giving us PQ KEM groups and PQ signature certs for the TLS layer.
OQSPROVIDER_REPO="${OQSPROVIDER_REPO:-https://github.com/open-quantum-safe/oqs-provider.git}"
OQSPROVIDER_REF="${OQSPROVIDER_REF:-0.9.0}"
# Identical optimization flags for every candidate (the credibility anchor).
# cortex-a76 is the RPi5 core. On non-A76 hosts (e.g. the macOS smoke box)
# setup.sh substitutes a safe fallback and records which flags were *actually*
# used in versions.lock + results JSON, so smoke numbers are never mistaken for
# the RPi5 baseline.
TARGET_CFLAGS_RPI5="${TARGET_CFLAGS_RPI5:--O3 -mcpu=cortex-a76}"
TARGET_CFLAGS_FALLBACK="${TARGET_CFLAGS_FALLBACK:--O3}"
# Pinned core for taskset on the RPi5 (4 cores: 0-3). Core 3 is chosen to stay
# away from CPU0 where the kernel tends to steer IRQs/RPS. Documented in README.
BENCH_CORE="${BENCH_CORE:-3}"