# Reset Policies Question: for explicit independent operations on private contexts, what does the reset between operations cost, and what does a generalized reset give up against one tuned to a fixed input? Three policies restore the 8 MiB arena of each context between operations: - `whole` copies the whole arena. - `pages` copies the pages that the first execution of a fixed input changed. It is valid only for unchanged code, input and host state. - `checkpoint` compares every page of current memory with the initial copy and rewrites the pages that differ, so it learns nothing from an earlier run. Globals and table references are restored too. All three runs use the same 21 kernels, arms and parameters in one session per machine: 7 paired rounds, 4 repeats, 16 jobs per round, native, conventional workers and `TaskPool` at widths 1, 2, 4 and 8, each batched and unbatched. Every sample checks the final context state and every return value. These are explicit independent operations, not automatic extraction. ## Results Paired median speedup over native for eight batched `TaskPool` workers, with process CPU as a ratio of native CPU. The full set is in the summaries. | Kernel | Machine | `whole` | `pages` | `checkpoint` | | --- | --- | ---: | ---: | ---: | | AES-CBC decrypt | Ryzen | 3.19x / 1.97 | 4.80x / 1.05 | 4.52x / 1.30 | | SHA-256 | Ryzen | 0.92x / 5.93 | 3.79x / 1.10 | 2.52x / 2.35 | | LZ4 compress | Ryzen | 0.44x / 12.67 | 3.11x / 1.52 | 1.14x / 4.59 | | xxHash32 | Ryzen | 0.43x / 12.62 | 2.48x / 1.80 | 1.40x / 3.93 | | AES-CBC decrypt | M2 Pro | 5.22x / 1.30 | 4.85x / 1.12 | 5.21x / 1.40 | | SHA-256 | M2 Pro | 3.69x / 1.66 | 5.15x / 1.11 | 3.93x / 1.81 | | LZ4 compress | M2 Pro | 2.06x / 2.80 | 3.07x / 1.64 | 2.27x / 3.05 | | xxHash32 | M2 Pro | 1.69x / 4.21 | 2.27x / 1.99 | 2.20x / 3.05 | On the Ryzen machine `pages` is ahead of `checkpoint` and `checkpoint` is ahead of `whole` for all 21 kernels. The checkpoint recovers most of the gain for large kernels and little for small ones, because it reads every page on every restore. On the M2 Pro the gaps are smaller and the ordering is not consistent: `whole` and `checkpoint` match `pages` for AES-CBC decrypt (5.22x and 5.21x against 4.85x), and `pages` leads on most other kernels. The earlier [changed-page study](../page-reset/README.md) reports 7.42x for AES-CBC decrypt with the `pages` policy under the same parameters on the same machine. This run reports 4.80x. Native (69.9 ms against 69.5 ms) and one-worker arms agree within 1% between the two runs. The eight-worker arm does not (9.44 ms against 14.55 ms), and the box load average was 0.1 at the start of the first run and 1.41 at the start of this one. The eight-worker figure for this kernel is a range from 4.8x to 7.4x on this machine, not a point value. ## Reproduction From the repository root with Node 26 and the built corpus (`node scripts/research/corpus.ts build build/research/corpus/build.jsonl`): ```sh FOLIO_CORPUS_RESET=checkpoint node scripts/research/corpus-mechanisms.ts build/research/reset-checkpoint 7 4 node scripts/research/analyze.ts build/research/reset-checkpoint/raw.jsonl summary.json ``` Use `whole` and `pages` for the other policies and a new output path for each run. `scripts/research/matched.sh` runs the three policies back to back under a shared timing lock with `nice -n 19`, `--max-old-space-size=128` and, on Linux, `taskset -c 4-11`. The Ryzen records come from that script on Linux with Node 26.10.0, cores 4 to 11 and the `powersave` governor. The M2 Pro records ran on macOS with Node 26.8.2 and no pinning. ## Artifacts - `ryzen-whole.jsonl`, `ryzen-pages.jsonl`, `ryzen-checkpoint.jsonl`: raw records, 2,499 samples each. - `m2-whole.jsonl`, `m2-pages.jsonl`, `m2-checkpoint.jsonl`: the same on the second machine. - `*.summary.json`: paired summaries generated from the matching raw file with `analyze.ts`. - `source.tar.gz`: the source the Ryzen runs used. The M2 Pro records name their sources by SHA256 in the environment row; the corpus harness and `src/` match the archive byte for byte. - `manifest.json`: byte lengths and SHA256 digests. Intervals in the summaries are bootstrap intervals over paired rounds within one run. They do not cover run-to-run variation, which the eight-worker arm shows is large.