Skip to content
vcfkit

Benchmarks

Results

Section titled “Results”
Operation vcfkit bcftools Speedup
filter -e 'INFO/AF < 0.01' 422 ms 1,695 ms 4.0×
normalize --fast --no-split 682 ms 2,820 ms 4.1×
normalize (standard, with noodles) 6,481 ms 2,820 ms 0.43×
liftover 6,713 ms — ~164K rec/s

1000 Genomes chr22, 1,103,547 variants · macOS aarch64 · bcftools 1.23.1 · full methodology →

Input

Section titled “Input”
  • Dataset: 1000 Genomes Project chr22 sites, extracted from the Phase 3 genotype VCF
  • URL: https://ftp.1000genomes.ebi.ac.uk/vol1/ftp/release/20130502/ALL.chr22.phase3_shapeit2_mvncall_integrated_v5b.20130502.genotypes.vcf.gz
  • Records: 1,103,547 variants
  • Format: plain VCF (uncompressed, no BGZ indexing)
  • Variants: predominantly biallelic SNPs + ~875 multi-allelic indel sites

Environment

Section titled “Environment”
  • Hardware: Apple M2 Pro (macOS aarch64)
  • OS: macOS 14
  • bcftools: 1.23.1
  • Rust toolchain: stable 1.75
  • Measured: 2026-04-19

Tool

Section titled “Tool”

hyperfine with 3 warm-up runs and 10 timed runs.

Commands measured

Section titled “Commands measured”
Terminal window
# filter
hyperfine \
  'vcfkit filter -e "INFO/AF < 0.01" chr22_sites.vcf' \
  'bcftools view -i "INFO/AF < 0.01" chr22_sites.vcf' \
  --warmup 3 --runs 10


# normalize --fast
hyperfine \
  'vcfkit normalize --fast --no-split -f hg19.fa chr22_sites.vcf' \
  'bcftools norm --no-version -c w -f hg19.fa chr22_sites.vcf' \
  --warmup 3 --runs 10


# normalize (standard)
hyperfine \
  'vcfkit normalize -f hg19.fa chr22_sites.vcf' \
  --warmup 3 --runs 10


# liftover (no bcftools comparison — plugin unavailable)
hyperfine \
  'vcfkit liftover -s hg19.fa -t hg38.fa -c hg19ToHg38.over.chain.gz chr22_sites.vcf' \
  --warmup 3 --runs 5

Notes on standard normalize

Section titled “Notes on standard normalize”

Standard normalize (without --fast) is 2.3× slower than bcftools. This is expected: the fast path is a raw-line loop; the standard path parses every record through noodles, which is a pure-Rust parser without htslib’s C optimizations.

The fast path handles biallelic SNPs and MNPs. Multi-allelic records and indels fall back to the standard noodles path regardless of --fast.

On typical variant call VCFs (mostly biallelic SNPs), --fast applies to the vast majority of records and the 4× speedup is representative.

Reproduce

Section titled “Reproduce”
Terminal window
# Download input data
mkdir -p tests/real_world
cd tests/real_world


# Download chr22 genotypes + extract sites-only
URL="https://ftp.1000genomes.ebi.ac.uk/vol1/ftp/release/20130502/ALL.chr22.phase3_shapeit2_mvncall_integrated_v5b.20130502.genotypes.vcf.gz"
wget -q "$URL" -O chr22_genotypes.vcf.gz
bcftools view -G chr22_genotypes.vcf.gz -O v -o chr22_sites.vcf


# Download reference FASTA
wget -q "https://ftp.1000genomes.ebi.ac.uk/vol1/ftp/technical/reference/human_g1k_v37.fasta.gz" \
  -O human_g1k_v37.fasta.gz
bgzip -d human_g1k_v37.fasta.gz
samtools faidx human_g1k_v37.fasta


cd ../..


# Run benchmarks
cargo build --release
hyperfine \
  './target/release/vcfkit filter -e "INFO/AF < 0.01" tests/real_world/chr22_sites.vcf' \
  'bcftools view -i "INFO/AF < 0.01" tests/real_world/chr22_sites.vcf' \
  --warmup 3 --runs 10