h-2 / bgzf_bench

Benchmark of BGZF implementations

This repository provides a small benchmark of different library implementations of the blocked GZip compression format.

Decoding

Setup

• Input file: ALL.chr20.phase3_shapeit2_mvncall_integrated_v5a.20130502.genotypes.vcf.gz
• Hardware: Thinkpad T14s, AMD Ryzen 7 PRO 4750U, WDC SN730
• The file is given as argument and the output is piped to /dev/null.
• Results are median out of three.
• "SeqAn2 4x" is the seqan2_4threads binary.
• "b.i.o. 4x" is the bio binary with 4 as second argument.

Results

Implementation	wallclock time	max RSS	% CPU
htslib	24.0s	3.5MB	99%
ShrinkWrap	22.7s	2.9MB	99%
SeqAn2	23.6s	3.8MB	112%
b.i.o	22.6s	5.3MB	99%
SeqAn2 4x	8.1s	5.9MB	316%
b.i.o. 4x	8.1s	6.8MB	317%

Observations

• The single-threaded implementations are quite similar in performance.
• C++ "iostream" based implementations are not slower than htslib!
• SeqAn2 single-threaded spawns two threads and thus uses more CPU than requested.
• The multi-threaded options provide noticeable speed-up.
• The multi-threaded options can't saturate their four threads. Using 5 threads instead of 4, reduces runtime to 6.3s and increases CPU usage to 417%.

Encoding

TODO