# Preparation performance measurements These measurements establish preparation and serialization costs, not comparative whole-system speed or performance budgets. The static fixture demonstrates negligible HTML size overhead. Regenerate with `node tools/benchmarks/measure.ts`. This rewrites this report and retains raw samples and correctness snapshots in `benchmarks/latest/` (ignored). Profile separately with `node tools/benchmarks/measure.ts --profile`. ## Conditions - date: 2026-10-09T12:37:35.285Z - node: v24.21.0 - cpu: 13th Gen Intel(R) Core(TM) i7-13620H - logicalCpus: 16 - memoryMiB: 31701 - warmups: 5 - repeats: 30 - shortRepeats: 100 - percentile: nearest-rank - gzip: zlib default - device: development machine, no pinned CPU or device profile - Warmups are excluded. Each workload records its own count; cold processes have no warmups. - p50/p95 use nearest rank; variance is sample variance in squared units. Byte rows are deterministic single observations. - Corpus throughput at median elapsed time: 5953.7 documents/second. ## Results | Workload | Samples | p50 | p95 | Variance | Boundary | | ------------------------------------- | ------- | ---------------- | ---------------- | -------- | ------------------------------------------------------------------------------------------------ | | static-100 prepare reused registry | 30 | 11.662 ms | 13.836 ms | 1.211 | Input and registry constructed outside timing; uncached preparation. | | static-100 input clone | 100 | 0.135 ms | 0.172 ms | 0.000 | Cloning only. | | static-1000 prepare reused registry | 30 | 113.095 ms | 130.156 ms | 42.403 | Input and registry constructed outside timing; uncached preparation. | | static-1000 input clone | 100 | 1.329 ms | 1.389 ms | 0.007 | Cloning only. | | static-10000 prepare reused registry | 30 | 1244.365 ms | 1272.924 ms | 1722.728 | Input and registry constructed outside timing; uncached preparation. | | static-10000 input clone | 100 | 14.299 ms | 16.394 ms | 1.069 | Cloning only. | | registry construction | 100 | 49.380 ms | 52.847 ms | 8.744 | Includes declaration checks and registry identity. | | static-1000 prepare new registry | 30 | 155.313 ms | 159.733 ms | 5.082 | Registry construction plus uncached preparation. | | static-1000 repeated same input | 30 | 109.569 ms | 112.956 ms | 2.103 | Repeated uncached preparation; no memoization. | | static-1000 fixture construction | 100 | 0.648 ms | 0.699 ms | 0.145 | Authoring helpers and fixture creation only. | | static-1000 end-to-end | 30 | 160.527 ms | 176.903 ms | 35.423 | Fixture construction, cloning, registry construction, preparation and serialization. | | static-1000 publication serialization | 100 | 1.567 ms | 1.761 ms | 0.014 | Already prepared document. | | publisle static html | 1 | 51946.000 bytes | 51946.000 bytes | 0.000 | Publication HTML only. | | publisle static html gzip | 1 | 2746.000 bytes | 2746.000 bytes | 0.000 | Default zlib compression. | | matched static html | 1 | 51927.000 bytes | 51927.000 bytes | 0.000 | Handwritten HTML for the same sentences; no competing compiler. | | matched static html gzip | 1 | 2732.000 bytes | 2732.000 bytes | 0.000 | Default zlib compression. | | matched mdx source | 1 | 52889.000 bytes | 52889.000 bytes | 0.000 | Source only; MDX compilation is unsupported. | | islands-5 prepare | 30 | 1.374 ms | 1.461 ms | 0.016 | Uncached preparation only; cloning, mounting and serialization excluded. | | islands-5 html | 1 | 2648.000 bytes | 2648.000 bytes | 0.000 | HTML only; excludes host bundles and network delivery. | | repeated-100 prepare | 30 | 22.620 ms | 23.811 ms | 0.237 | Uncached preparation only; cloning, mounting and serialization excluded. | | repeated-100 html | 1 | 51256.000 bytes | 51256.000 bytes | 0.000 | HTML only; excludes host bundles and network delivery. | | distinct-20 prepare | 30 | 3.353 ms | 3.693 ms | 0.012 | Uncached preparation only; cloning, mounting and serialization excluded. | | distinct-20 html | 1 | 6406.000 bytes | 6406.000 bytes | 0.000 | HTML only; excludes host bundles and network delivery. | | inline-code-100kb prepare | 30 | 2.602 ms | 3.009 ms | 0.063 | Uncached preparation only; cloning, mounting and serialization excluded. | | inline-code-100kb html | 1 | 100156.000 bytes | 100156.000 bytes | 0.000 | HTML only; excludes host bundles and network delivery. | | corpus-10000 prepare | 30 | 1679.635 ms | 1756.027 ms | 7441.658 | Sequential whole-corpus elapsed time; reused registry. | | contract-graph-1000 walk | 100 | 0.105 ms | 0.123 ms | 0.000 | Identity and edge checks only; no schema compilation. | | static-1000 process-cold prepare | 30 | 145.618 ms | 162.395 ms | 125.978 | First prepare in each fresh process; registry and fixture already constructed. | | static-1000 process start-to-exit | 30 | 361.464 ms | 391.323 ms | 491.635 | Parent elapsed time including process startup, imports, registry, fixture and first preparation. | ## Profiling findings The [preparation analysis](performance-analysis.md) records the optimization and its before/after evidence. Canonical JSON byte accounting now avoids buffer allocations for small emitted fragments and retains native encoding for large strings. Declaration validation, JSON bounds and identity semantics remain unchanged. ## Limits - Budgeted prepare rows stay uncached. A separate host compilation cache reuses unchanged corpus documents and is not a numerical cap. - HTML sizes exclude CSS, JavaScript, island props delivered separately, requests, activation latency and reader memory. - Repeated paragraphs compress unusually well; mixed-content size comparisons remain follow-up work. - Inline code text is not fetched dataset delivery. Island preparation does not measure browser mounting or selective bundle loading. - MDX compilation, other publishing toolchains and browser delivery measurements remain unsupported. - Timing variability reflects this machine and execution order; cold processes are reported separately from warmed workloads.