Benchmarks

runlength compared with Bioconductor's IRanges and S4Vectors, the standard run-length encoding in R, and with pyrle.

Results

Times in milliseconds; each is the best of three runs.

Operation	runlength (Python)	runlength (Rust)	IRanges (R)	pyrle
Coverage of 1M intervals	38	32	334	271
a + b	20	17	197	86
a < b	18	14	139	124
a.mean()	2	1	18	0.5¹
Sum over 1M slices	34	26	125	–
1M slices, copied out	94	90	–²	–
Coverage of 1M intervals, 25 keys	34	17	351	390
batch + batch, 25 keys	10	3	59	21

• The workload: 1M random intervals of length 100 on a sequence of length 100M, giving coverage with about 2M runs (a); a second, independent coverage (b); 1M slices of length 1000; and for the keyed rows, the same intervals spread over 25 keys. a + b has about 3.9M runs.
• Machine: Apple M3 Pro (12 cores), runlength from master (after 0.1.0), R 4.5.0 with IRanges 2.42.0, pyrle 0.0.41 on Python 3.12.
• runlength (Rust) is the core crate with i64 values and integer keys. The Python package goes through the Arrow layer, which handles any value type. For numbers without nulls its operators use the same single-pass loop as the core, so the two are close; coverage by key also hashes the string keys, which come from a pandas DataFrame column.
• ¹ pyrle's mean() is quick because it ignores run lengths, which makes it wrong: see Coming from pyrle.
• ² IRanges doesn't copy slices: converting Views to an RleList takes 0 ms because the list still points into the original. The closest comparison is the sum over slices (viewSums), which does visit the data.
• pyrle has no batch slicing; "–" means the library has no matching operation.

Running them

cargo run --release --example bench                    # runlength (Rust)
python scripts/bench.py                                # runlength (Python), and pyrle if importable
PYTHONPATH=~/code/pyrle python3.12 scripts/bench.py    # with pyrle built in place
Rscript scripts/bench_r.R                              # IRanges

The R script needs the Bioconductor packages S4Vectors, IRanges and GenomicRanges.