Reader behavior under measurement.

What this page records

The baseline used 3 measured runs after 1 warmup per workload, with an isolated process for each run. It contains 13 representative measurements and 30 correctness-only measurements; correctness-only rows verify reader contracts without claiming representative large-input performance.

Recorded on Aug 15, 2026 from commit 6acf53304b44 using Intel(R) Core(TM) i7-10700 CPU @ 2.90GHz with 16 logical CPUs. The benchmark process reported Linux 6.17.0-41-generic.

Measurement boundary

Each workload opens a scientific source through the same public reader boundary, detects and opens its document, enumerates datasets, opens the selected dataset, reads the selected plane or series, validates the result, and closes all resources. The report keeps process startup, module import, registry construction, fixture preparation, and validation evidence separate from the core operation timing.

• Timing: total wall time, time to the first emitted block, and complete selected operation are reported as median / p95.
• Source behavior: read calls, requested and returned bytes, unique source bytes touched, overfetch, and companion resolution are measured.
• Memory: absolute process peak RSS is accompanied by external memory, ArrayBuffer memory, output bytes, and the largest emitted block.
• Correctness: output sample type, component count, selected sample count, block count, and format-specific calibration assertions are checked before a result is published.

Reader coverage

Reader-level values below are the maximum median observed among that reader's baseline workloads. They are useful for finding source and memory boundaries, not for ranking unrelated formats against one another.

Baseline workload detail

Timing and source columns show median / p95 where both are available. Peak RSS is the median absolute process peak. The workload title and support boundary describe the fixture-backed claim; they do not imply a universal guarantee for every file in that format.

Range-backed coverage

The separate range profile covers 24 direct-range workloads across 19 readers at 0, 5, 25, 100 ms of underlying-read latency. It is designed to expose round-trip sensitivity and coalescing behavior without pretending that a full-file reader has remote range semantics.

Nanonis SXM, Igor Binary Wave v5, X3P surface exchange, ENVI, FITS, MRC/CCP4, TIFF, OME-TIFF, Aperio SVS, Gatan DigitalMicrograph, FEI/Thermo TIA SER, FEI/Thermo TIA EMI, NCEM EMD 0.2, FEI/Thermo Velox EMD, Lispix RPL/RAW, MetaImage MHD/MHA, NumPy NPY, NanoMegas ASTAR blockfile, Quantum Detectors Merlin MIB.

The published snapshot above is the zero-latency baseline. Run the range profile locally for the expanded latency matrix and inspect its isolated-process report before making performance claims.

Reproduce the evidence

The harness and fixture policy are documented in the repository's scientific reader source and memory harness section.

npm run bench:scientific:smoke
npm run bench:scientific:baseline
npm run bench:scientific:range
npm run bench:scientific:full

Use smoke for a fast contract check, baseline for the page's workload set, range for direct-range latency coverage, and full for the complete profile. Generated fixtures and raw reports stay under the ignored benchmark artifact directory; the page's JSON snapshot is the reviewable public surface.