Measures the JCA cryptographic primitives a JVM or an Android device provides: what is registered, what runs, and how fast.
pip install -r requirements.txt
cp config/facade/settings.example.yaml config/facade/settings.yaml
cp config/perf-repair/perfrepair.example.yaml config/perf-repair/perfrepair.yaml # set apps.targets
python3 apps/cli/facade.py --settings config/facade/settings.yaml --check # the resolved tree; no app build, device or model
python3 apps/cli/facade.py --settings config/facade/settings.yaml # benchmark, then guided repair
| read | for |
|---|---|
| docs/facade.md | the settings index, its direction rule, the phases |
| docs/run.md | prerequisites, apps/benchmark-runner/pipeline.py flags, device runs |
| docs/architecture.md | modules, dependencies, stage files |
| docs/results.md | results/<device_id>/ layout and manifest.yaml |
| docs/cryptography/index.md | configured primitives and providers |
| docs/index.md | every cross-module doc |
| apps/perf-repair/README.md | crypto replacements worth testing in an app |
- Set
run.thermalGateon a phone, and find its temperature for your device. A hot phone slows its CPU without reporting it, so numbers shift silently.temperatureKeptCelsius36.0 is where a Pixel 10 slows down; other devices may differ. - Give very short algorithms more runs. A case contains runs and a run contains iters. Fast hashes and MACs vary more from run to run than slow operations do, and more iters do not fix that: the extra launches do.
run.processRepetitionssets one count for every case.
- Turn on airplane mode.
- Keep the charger connected for the whole run.
- Start with the battery at 95% or more.
- Enable USB debugging and connect over USB.
- Close every other instance of the benchmark app.
- Do not touch or move the phone while a run measures.
These are not settings.
- A full GC is forced between iters, outside the timer; GC pauses are reported beside the time.
- The app and every bundled provider are compiled ahead of time (
speed) before measuring. - Energy is measured in its own pass; timing is never traced.
- Energy windows are at least 5 s (
run.passes.energy.minWindowSeconds). - Each process prepares and measures one case.
- Each process runs its cases in an order shuffled from
run.orderSeed; an algorithm's fixed and notFixed cases run back to back. - A failed launch is retried
run.launchRetriestimes. - A launch with too-short iters after warm-up is measured once more, then recorded as failed.
- A run with fewer samples than asked is relaunched.
- A failed attempt is kept in
rejected/with its reason. - A stopped run resumes from its finished blocks.
- One run at a time per phone.
- Every skipped case is recorded with its code in
skipped.yaml.
- Real apps start JIT- or profile-compiled; these numbers are ahead-of-time compiled.
- GC cost is reported beside the time, not in it.
- The measuring thread stays on one big core.
- Clocks are not locked.
- The phone is assumed not to throttle; temperature is recorded per case and held only with
run.thermalGate. - Each launch generates its own key.