Microscopy Pareto
Find the speed and compression tradeoffs for your machine, then compare hosts on the same image inputs.
Loading measurements…
Previews use adjusted display contrast.
Sampled pixel entropy
Pixel entropy is Shannon entropy of complete stored pixel values, in bits per pixel. Each distinct float32 bit pattern counts as one value.
Samples use eight evenly spaced full rows from every plane, before shuffle or padding. “Sample limited” marks samples with every pixel distinct: the value reaches log₂(pixels sampled), leaving the dataset’s entropy unresolved. These histograms omit spatial correlations and do not determine achievable compression. Sample metadata includes row indices and checksums.
| Dataset | Type | Pixel entropy (bits/pixel) | Distinct values | Pixels sampled |
|---|
Connected points form the observed frontier. Bars show min–max observations, not confidence intervals. Blosc uses bitshuffle.
No configurations match these filters.
Configuration measurements
| Setting | Backend / sink | Chunk | Block | Logical GiB/s | Fold × | Evidence | Run date (UTC) |
|---|
How to read these results and source data
Throughput counts logical image bytes per second through final drain and close. Compression fold divides logical input bytes by physical output bytes. Padding can make even uncompressed output larger than the logical image, giving a fold below 1. The horizontal axis is logarithmic: equal distances represent equal ratios.
A setting is on the observed frontier when no other tested setting in the same dataset, machine, backend, and sink is at least as good at both objectives and better at one. The All datasets view overlays inputs on common axes within each panel, with one frontier per dataset. Filtering recomputes those frontiers. Raw controls participate in the same comparison. Connections join measured settings and do not predict performance between them. Frontier view includes the complete observed ranges of frontier settings; All points restores the full comparison.
Each point summarizes its own observations. Ranges describe the recorded minimum and maximum; they are not confidence intervals. Two or three observations cannot establish a precise probability of frontier membership. Reference runs help characterize variation across a session, but these observations do not establish variation between sessions. Different sessions and reference runs are not pooled.
CPU and GPU measurements compare complete machine configurations. Coverage differs across datasets and backends; configuration and observation counts appear on each plot. These measurements do not establish an exhaustive frontier for every codec or machine.
Blosc blocks subdivide compression within each chunk. Requested block sizes are recorded explicitly; the compressor may adjust its actual layout. Raw CPU LZ4 uses the LZ4 HC encoder. Comparing raw LZ4 with Blosc-LZ4 includes differences in the encoder, shuffle, and block size.
Filesystem throughput includes platform buffering, final drain, and close. Pipeline stage intervals overlap, and each stage uses its own input/output bytes: stage times and rates do not sum to end-to-end performance. Exact replay geometry, shard counts, source revisions, and raw observations are available in configuration details. These analyses are separate from the regular performance sweeps in Over time.
When at least six matched rounds are available, additional rings mark settings on the frontier in at least 5% of whole-round resamples. Resampling includes all selected settings in the source condition, even settings hidden by filters. Approximate 95% bootstrap intervals appear in details. This frequency is resampling support, not a posterior probability or a simultaneous confidence guarantee.
Dataset previews and credits
These are the thumbnails supplied with the measured inputs. Display contrast was adjusted for visibility; the raw benchmark pixels were not rescaled.