All comparisons were run on a set of auto-generated FASTA or FASTQ sequences with lengths normally distributed around 300 bp and loaded into memory. For FASTA, sequences were either on a single line ("FASTA" in barplots) or wrapped at the length of 80 (4 sequence lines).
The bars represent the throughput in GB/s (+/- standard error of the mean), as measured using criterion. Run on a Intel i7-8550U with turbo boost disabled.
To run the benchmarks and generate the summary charts, run
scripts/run_bench.sh. The charts need R and ggplot2 installed.
Different aspects of the API were compared, always using the same input for a given format. The darker bars show results for parsers from other crates (fastq-rs and Rust-Bio).
A few points on the results:
- The standard way of iterating is fastest, any operation involving
allocations (such as
RefRecord::to_owned_record()) slows down the process. - FASTX readers are almost as fast as readers specialized on one format.
The
seq_io::fastx::dynamicreaders were slower thanseq_io::fastxin this benchmark. - Access to the contiguous sequence is fast if the sequence consists of a single
line. If wrapped to multiple lines,
RefRecord::full_seq_givenis the best solution. Similarly and not surprisingly,RefRecord::clone_into_owned()is the fastest solution for obtaining anOwnedRecordfromRefRecord. - Accessing the sequence ID has some performance impact, since the space
separating ID and description needs to be searched. However, checks for
validity of
UTF-8had the strongest impact. - A specialized reader only accepting single FASTA lines (
seq_io::fasta::single_line) is considerably faster.
Readers were also tested with different internal buffer capacities using randomly generated single-line FASTA and FASTQ input with different sequence lengths.
The current default buffer capacity is 64 KiB.
