Viewer performance
Current production browser
The production /demo route now uses the deterministic, anchored SVG tube-map renderer described in ADR 0003. Its current built-browser evidence, bundle comparison, screenshots, and configured-archive measurements are retained in results/2026-08-27-simple-browser-v1/REPORT.md.
In the retained Chromium observation, HLA-B completed in 1,431.9 ms after opening the configured 8.2 GiB immutable archive, with the first tile visible at 1,254.6 ms. The renderer laid out 141 node groups, 240 topology edges, and eight weighted tile-local patterns in 6.5 ms. MICB completed in 664.1 ms on the already-open archive with 82 groups, 153 edges, and eight patterns. These are single public-network observations, not latency distributions.
The remainder of this page preserves the earlier Canvas-explorer measurements because they remain useful reader/decode evidence. They are not measurements of the current production UI.
Measurement model
Elapsed wall time and aggregate task work are different quantities. Reader traces merge concurrent time intervals before reporting a wall-clock phase. Integrity and regional-decode intervals are also merged, so concurrent tiles are not added together and mislabeled as elapsed time.
decompressionMs is decompression critical-path wall time. decompressionTaskMs is the sum of individual decompression task durations and may exceed wall time for an asynchronous/parallel decompressor. The benchmark wrapper preserves synchronous decoder returns; this prevents time waiting for unrelated decode continuations from being attributed to decompression.
The retired explorer added object-open, first-summary-paint, first-tile-paint, query-complete, viewer-model, layout, paint, and rolling frame-p95 evidence. The range waterfall retains exact offsets, lengths, and logical layers. Browser origin logs remain authoritative for actual request counts and bytes; reader plans and cache hits are reported separately.
Retained 100 kb detailed query
Workload: CHM13#chr1:1,000,000-1,100,000, context 100, Chromium, loopback strict-range origin, 8,832,749,949-byte current-v1 archive SHA-256 ecf5ae4fa8c784a80307507f58bed894311b8560724b57de0fcc35237c324b63. Every result matched canonical hash b191be02fc2a9556349d8b5b97b268c90c579b1c275cc600355bfaae5b499473.
| Metric | Retained before | Current pure JS | Current WASM |
|---|---|---|---|
| Cold no-store query | 1,678.3 ms | 1,565.3 ms | 1,458.1 ms |
| Cold total including open | 1,717.3 ms complete UI | 1,577.7 ms harness | 1,485.3 ms harness |
| Actual requests / bytes | 4 / 294,198 B | 4 / 294,358 B | 4 / 294,358 B |
| Decompression wall/task | previously overlapping | 32.2 / 32.2 ms | 6.5 / 6.5 ms |
| Integrity | not exclusive | 32.8 ms | 32.8 ms |
| Regional decode | 1,352.8 ms accumulated | 1,373.4 ms wall union | 1,288.5 ms wall union |
| Merge | not comparable | 55.5 ms | 63.4 ms |
The pure-JS cold query improved 6.7%, and the WASM observation improved 13.1% against the retained query wall. Neither reaches the requested 2x target (<=839.2 ms). The measurements show that zstd is no longer the governing phase: WASM reduced decompression by 25.7 ms but reduced the observed query by only 107.2 ms relative to pure JS in this run. Regional payload decode and GBWT reconstruction dominate.
WASM adds 251,806 deployable bytes and measured 6.8 ms one-time initialization. It therefore remains optional. Bounded decompression/reconstruction workers, transferable-buffer reconstruction, and OffscreenCanvas were not adopted: there is not yet a controlled implementation that demonstrates an end-to-end or interaction win without extra copies and lifecycle complexity.
Search, summaries, and public origin
One public-path observation after archive open measured exact HLA-B at 125.5 ms cold (69,112 bytes, two dependency rounds) and 1.2 ms warm (zero bytes). A limited HLA- prefix returned five results with an explicit truncation flag from cached pages. The 100 kb CHM13 summary measured 106.2 ms cold (49,503 bytes, two rounds) and 0.3 ms warm. These public-network observations are not relabeled as loopback percentiles.
The origin probe passed HEAD, exact and overlapping/tail 206 reads, byte comparison with the checksum-matched local object, stable ETag, identity encoding, immutable/no-transform caching, CORS, exposed headers, and preflight for the GitHub Pages origin. It transferred only the probed ranges.
Cache and browser qualifications
- Cold library/normal HTTP cache uses a fresh Playwright context; operating system caches are not controlled.
- Warm-library scenarios force transport
no-storeso library and browser HTTP caches are not conflated. - The 100 kb comparison is loopback runtime evidence, not CDN latency.
- Public search/summary measurements are a single observation, not a p50/p95 corpus.
- Heap evidence is Chromium's exposed JavaScript heap only and excludes native and WASM memory.
- Frame p95 is a bounded canvas/app sample; a full DevTools long-task and Core Web Vitals trace remains future evidence.
Raw requests, scenarios, phase measurements, environment, screenshots, and the origin report are retained in results/2026-08-27-viewer-explorer-v1/.