Microscopy Pareto
Find useful tradeoffs across chunk size, Blosc block size, and codec. Compare CPU, GPU, and sinks on the same image input.
Loading retained studies…
Blosc uses bitshuffle. Raw LZ4/Zstd and uncompressed controls compete on the frontier and remain visible when filtering Blosc blocks.
Prominent connected markers form the observed frontier; faint markers show other candidates. Lines show observed min–max ranges: throughput vertically and compression fold horizontally when available. Each panel has its own frontier. Configuration details report reference variation.
No configurations match these filters.
Configurations
| Setting | Backend / sink | Chunk | Block | Logical GiB/s | Fold × | Evidence |
|---|
Reading the study
Throughput counts logical image bytes per second through final 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 (base 10): equal distances represent equal ratios. Positive folds below 1 remain on the plot; nonpositive or unavailable objectives are reported as unavailable and remain in the table.
A configuration is on the observed frontier when no other tested setting in the same study, image input, backend, and sink is at least as good at both objectives and better at one. Filtering recomputes that frontier among the remaining settings. Raw controls participate in the same comparison. Connections join measured settings and do not predict performance between them.
One observation leaves run variation unknown. Observed min–max ranges describe the recorded runs; they are not confidence intervals. Comparison studies keep varying conditions on the observed frontier and resample whole matched rounds to describe uncertainty. The retained discovery keeps its original reference-drift exclusions. Pilot results are labelled separately and estimate execution cost.
For comparison studies, rings highlight additional settings that appear on the frontier in at least 5% of round resamples. This frequency uses every selected setting in the study, image, backend, and sink condition, including settings hidden by filters. Approximate 95% bootstrap intervals appear in configuration details. Resampling support is not a posterior probability or a simultaneous confidence guarantee; one session does not characterize variation between sessions.
Blosc blocks subdivide compression within each chunk. Requested block sizes are recorded explicitly; the compressor may adjust its actual layout. CPU raw LZ4 uses an HC entry point, whereas CPU Blosc-LZ4 uses LZ4, so their difference includes more than framing.
Pipeline stage intervals overlap. Their times and rates help explain limits but do not sum to elapsed time. Study measurements are separate from the regular hero and regression sweeps in Over time.