perf(pipeline): size alignment batches so the result vector stays off the arena path

[BenjaminDEMAILLE]

Independent of #256 / #257 — this touches only src/lib.rs and can merge in any order.

What

run_batch_pipeline allocates a fresh Vec<Result<T, Error>> per batch and drops it once consumed. At the fixed 10,000 reads that vector is 1.7 MB, past mimalloc's large-object threshold, so every batch was served from the arena path rather than a thread-local page: a fresh mapping, and — with arena_eager_commit off, which main sets deliberately to keep RSS down on genome-scale runs — every page faulted back in.

Profiling a 2M-read solo run put ~38% of samples under mi_huge_page_alloc / _mi_arenas_page_alloc, against 4,370 for the alignment itself. The allocator was costing nearly as much as the work.

batch_size_for::<T>() derives the count from the element size and caps it at a measured ceiling.

The ceiling is measured

Sweeping batch size on the solo run, 8 threads:

batch	10000	3000	2000	1500	1000	500	250
wall	3.46s	2.29s	2.15s	2.13s	2.17s	2.35s	2.58s

Plateau from ~1000 to 2000; below it per-batch dispatch starts to dominate, which is what the floor guards.

Sizing purely by bytes was not enough on its own: some per-batch cost tracks read count rather than the result vector, so the byte rule alone picked 4096 for the solo product and left time on the table (2.47s vs 2.14s). 4096 also cost ~1.5% on paired-end, which 2048 does not. Hence both a byte target and a count cap. The byte rule still earns its place — it keeps the allocation under the threshold if a result struct grows later, rather than leaving a tuned constant to rot.

Measured

Apple M4 Max. Solo workload is 2M reads, 5000 barcodes, 4546 genes, CB_UMI_Simple.

workload	before	after	change
solo, wall, 8 threads	3.49s	2.15s	-38.4%
solo, user CPU, 1 thread	15.03s	10.63s	-29.3%
yeast PE, user CPU, 1 thread	18.72s	18.68s	neutral
nfcore PE, user CPU, 1 thread	25.44s	25.48s	neutral

Paired-end is unchanged because its per-read work already amortizes the batch allocation. The win lands on workloads whose per-read cost is small, which is what a clean solo run looks like — and, I would expect, a production 10x run.

Correctness

Byte-identical: solo Aligned.out.sam, matrix.mtx, barcodes.tsv, features.tsv; and Aligned.out.sam + SJ.out.tab on both paired-end sets.

• 593 tests pass
• 0 clippy warnings (--all-targets --release)
• cargo fmt --check clean

Note on the benchmark

There is no solo dataset in the repo big enough to measure (the checked-in one is 400 reads), so I generated one: a synthetic gene model over the yeast genome (4546 genes, 1-3 exons each), a 5000-barcode whitelist, and 2M reads with 2% single-base barcode errors so CB correction is actually exercised. 95.2% uniquely mapped. Happy to contribute the generator if a solo benchmark fixture would be useful to have.

🤖 Generated with Claude Code