How much does a small read cost?

Compare density and resident read cost. This is a historical genome example, not a prediction for your files.

lzbench and TurboBench rank compressors by density and bulk throughput; SeqBench covers sequence compression; per-format seekable readers exist for gzip (rapidgzip), zstd (zstdra, seekable-zstd) and BGZF (htslib bgzip). None of them compares the cost of reading one region across formats at matched block sizes. That is what this measures.

Matched-g: interactive ACEAPEX vs BGZF

Measured on thirty frozen 2 MiB T2T-CHM13v2.0 windows, not the chr1 snapshot above; ratios are not comparable between the two tables.

ACEAPEX is written by the author of this benchmark. Its density advantage and its latency penalty are both reported below; the harness, corpora and raw samples are in the repository.

ACEAPEX interactive: LIT_CHUNK=64 KiB, FSE_CHUNK=4 KiB. Both codecs use the same runner, frozen corpus and resident 16 KiB request trace.

Matched-g interactive report · GitHub Actions run 35364484648

Independence cost at 16 KiB (separate chr1 c_file(g) scope): ACEAPEX default @ ee5a37e costs 1.632% versus 6.569% for matched zstd-seekable; BGZF is n/a because no same-encoder whole-input baseline exists. Scope and provenance.

Batch sweep scope (2026-09-19): hg38 chr1, N=5000, 16 KiB ranges, workers=1, Actions run 35427975099.

A native batch API is what makes the dense profile usable at all for region reads: it moves from 659.6 to 12,369.7 ranges/s under uniform access. But dense leads interactive only under uniform access. Once requests concentrate, interactive overtakes it somewhere between 12 and 8 bits of access entropy and keeps the lead: at Hα ≈ 2 bits, interactive reaches 869,088 ranges/s against 467,998 for dense. For batch-heavy workloads with any locality, interactive is the profile to pick.

Concentration helps BGZF more than it helps ACEAPEX in the single-request path. Under uniform access bgzip runs 10,040.6 ranges/s against ACEAPEX interactive at 9,086.4, a 1.1× gap. At Hα ≈ 2 bits bgzip reaches 64,722.7 against 13,847.4, a 4.7× gap. The batch API is where ACEAPEX answers locality; the loop path is not.

Access-entropy sweep.

Try your own data · What is released and what is measured?

The slider selects five measured profiles; it never interpolates. Batch workers: 1; requests: 5,000.

Values behind the plot

Green marks a best observed value only where direction is meaningful and scope matches. Granularity and access break-even are descriptive properties, not rankings. Access entropy appears only in the batch table. Missing values are never plotted as zero.