Skip to content

Repository navigation

README.md

Latest commit

 

History

History
57 lines (39 loc) · 3.1 KB

README.md

File metadata and controls

57 lines (39 loc) · 3.1 KB
Raw

Headless Subsecond spike

Measured on an Apple M2 Max build Mac. See the research report for conclusions and qualifications. evidence/ contains the recorded JSON results; results/ contains disposable full logs from a run.

This is a Rust terminal binary, with no Dioxus UI or window. main owns the state and runs a nominal 120 Hz loop. The patched function returns a string. The Python driver edits Rust source and consumes actual frame events, using timestamps from the same machine.

Run

Copy this directory into its own scratch Git repository on the build machine, outside the Rune workspace:

mkdir -p <scratch>/rune-hot-patch && git -C <scratch>/rune-hot-patch init
rsync -az --exclude target --exclude results --exclude evidence \
  spikes/hot-patch/ <scratch>/rune-hot-patch/

If Dioxus CLI is missing, install the release used by this spike:

cargo-binstall dioxus-cli --version 0.7.10 --no-confirm
"$HOME/.cargo/bin/dx" --version

Homebrew's dx is Deno's command, not Dioxus. The harness deliberately uses the absolute Dioxus path.

Run each mode from the scratch repository, one at a time:

for mode in builds benchmark probes extras; do
  python3 run.py "$mode"
done

builds requires a fresh artifact directory; choose a new RUNE_NORMAL_TARGET if target/normal already exists. This prevents labeling a warm build as cold. For example, put env RUNE_NORMAL_TARGET=target/normal-new before python3 in the command above. Each probe launches a fresh process, patches once, records its observations, stops the process group, and restores the source. Do not run modes concurrently.

The boundary-signature probe can be run alone:

python3 run.py extras --case explicit_hotfn_signature

The underlying headless command is:

"$HOME/.cargo/bin/dx" serve --hot-patch --platform server \
  --open false --interactive false --addr 127.0.0.1 --port 18743

The harness disables any inherited sccache wrapper, enables incremental compilation, and limits Cargo to two jobs. The measured runs used background scheduling priority, under which a nominal 120 Hz sleep loop runs in bursts; this is not a display-pacing benchmark.

Interpreting the probes

  • The benchmark uses a two-reference closure capture to work around 0.7.10's pointer-sized-closure dispatch bug. extras reproduces the broken single capture and verifies HotFn::current(render).call((&state,)) as a cleaner integration.
  • The layout probe adds a field but never explicitly reads it. It demonstrates that patches can cross incompatible layouts without a rejection or migration; it does not establish memory safety.
  • A successful revision with unchanged PID, increasing counter, unchanged state/Vec addresses, and vec0 == counter + 73 establishes preservation of this sample's main-owned state only.
  • Private and exported tip-crate statics, thread locals, and cached trait objects have separate observations. Their behavior must not be inferred from the main-owned-state result.
  • p95_nearest_rank_ms is the largest of ten samples. There is no statistical claim beyond this small run.