An auto tiler for macOS
parcel Benchmarks hosts.md
1.6 kB · 30 lines
Markdown
at main

Measurement hosts #

No remote host is selected or approved by default. Pass an explicit SSH host to Scripts/bench-remote.sh or Scripts/bench-host-setup.sh. Setup defaults to an offline dry run; --check reads remote configuration and --apply changes only the listed AC power settings and backup exclusion. No host was exercised while implementing these scripts.

Before measuring, use a dedicated console session, AC power, a matching Xcode and Swift toolchain, and a quiet window. Add ~/parcel-ci to Spotlight Privacy. Grant Accessibility to the host's build/Parcel.app manually and verify with PARCEL_E2E_PREFLIGHT_ONLY=1 Scripts/test-e2e.sh. Optional privileged energy and leak tools need an explicitly reviewed sudo policy, not unrestricted sudo.

Remote runs poll quietness for BENCH_WAIT=15m, record actual readings, and fail with exit 3 if the host never qualifies. Build and tool failures are not quiet failures and are not retried. Runs use caffeinate -i -s and nice -n -5; a priority-setting failure is surfaced, never replaced with another scheduling class. Retrieved reports live in build/remote/<host>/<UTC timestamp>/. Newly recorded baselines return unstaged for deliberate review.

Role Approved host Last baseline
Primary instruction gate Not selected None
Intel CPU-only cross-check Not selected None
Live smoothness / idle / soak Not selected None

Record an approved host and dates here only after real preparation and runs. Counter support, OS build, toolchain, host identity and quietness belong in each report; hardware model alone is not a host identity.