Runtime profile and allocation pass (2026-07-21)

Scope and reproducibility

The baseline is dev commit 9a036db. Profiling and timing used LongBench H69 PacBio name-collated BAMs, four threads, --coverage-model auto, --seq-tech pac-bio, --endpoint-weight 0.3, and --em-accel none. Disabling acceleration deliberately exposes the cost of repeated fixed-point evaluations. Raw profiler data is retained outside the repository at oarfish-evaluation-data/profiling-20260721/h69-250k.data; repeated benchmark tables are in runtime-opt-baseline-9a036db-20260721 and runtime-opt-pass1-20260721 under the same evaluation-data directory.

Baseline profile

On H69 250k, the median/mean run is about 5.0 seconds. Instrumented phase means from three baseline runs were:

Phase	Seconds	Share of 5.04 s wall time
Alignment parsing/filtering	0.794	15.8%
Coverage modeling, including anchor inference	2.007	39.8%
Final EM	1.597	31.7%
Digest, output, and other overhead	0.642	12.7%

The sampling profile attributes roughly 53% of CPU samples to the inlined EM fixed-point work. Other visible costs include zlib inflation (8.4%), Rayon scheduling/reduction (7.1%), memory clearing (4.2%), memory copies (2.5%), BAM decoding (1.4%), and the physical-endpoint likelihood path (about 1.4% including children). Thus the two inference passes, not endpoint fitting itself, are the main optimization target.

Retained changes

1. Parallel EM now writes each fixed-point result directly into the driver's destination buffer. The old path accumulated into a second transcript-sized vector and copied it after every evaluation (2,000 copies for the two unaccelerated PacBio auto inference passes in this benchmark).
2. Hybrid/adaptive coverage combination reuses one scratch vector across reads rather than allocating and freeing a vector for every read. PacBio auto calls this path twice, so H69 250k avoids about 400,000 short-lived allocations.
3. Physical-endpoint likelihoods are appended and normalized directly in their final packed vector rather than first allocating a temporary vector per read.
4. Logistic coverage normalization reserves its known final packed capacity.

The optimized and baseline H69 250k quantification files are byte-identical. All 34 release tests pass.

Five-run perf stat measurements show the retained pass reduces aggregate CPU task time from 14.110 to 13.482 seconds (-4.5%), cycles from 67.72 to 65.46 billion (-3.3%), and cache misses from 824.3 to 806.3 million (-2.2%). Wall time was scheduling-sensitive on the shared machine (5.30 +/- 0.09 seconds baseline versus 5.46 +/- 0.15 optimized in that ordered run), so no wall-clock claim is made from that noisy comparison. The smaller H69 50k three-run phase means did move in the expected direction: coverage 0.687 to 0.670 seconds (-2.5%) and EM 0.565 to 0.555 seconds (-1.7%), with essentially unchanged total wall time and RSS.

Rejected experiment

The raw-read producer and mapper batch vectors currently retain thread-local capacity but clone batches when sending them through channels. Moving the buffers instead removed those copies and reduced one 50k SIRV raw-read run from 7.62 to 7.47 seconds, but peak RSS increased from 106 to 120 MiB as replacement buffers churned through the allocator. Quantification differed only by 4.55e-13 reads (parallel floating-point order). This tradeoff was reverted. A future version should use an explicit buffer-return pool if this path is revisited, preserving both ownership transfer and thread-local capacity.

Next targets

The dominant remaining memory traffic is clearing every dense per-shard transcript accumulator and reducing all shards across all transcripts on every EM evaluation. A sparse touched-transcript scheme is promising only when the active set is small; it needs workload-sensitive selection because dense human transcriptomes can make bookkeeping slower than memset. BAM decompression is the next independent phase worth testing with parallel block decompression or a faster input representation.