BWA-FastAlign: Faster and Cheaper Sequence Alignment on Commercial CPUs

BWA-FastAlign is a high-performance, cost-efficient software package for mapping low-divergent sequences against a large reference genome, such as the human genome.

It is designed as a drop-in replacement for the de facto standard BWA-MEM, offering 2.27× ∼ 3.28× throughput speedup and 2.54× ∼ 5.65× cost reductions on standard CPU servers, while guaranteeing 100% identical output (SAM/BAM) to BWA-MEM.

🚀 Key Features

• High Throughput: Achieves ~2.85× average speedup over BWA-MEM by optimizing both the seeding and extension phases.
• Cost Efficient: Delivers 2.54× ∼ 5.65× cost reduction compared to state-of-the-art CPU and GPU baselines (including BWA-MEM2 and BWA-GPU).
• Identical Output: Guarantees 100% output compatibility with BWA-MEM. You can swap it into your existing pipelines without changing downstream analysis results.
• Low Memory Footprint: Uses a novel Multi-stage Seeding strategy (Hybrid Index) that improves search performance without the massive memory overhead seen in hash-based or learned-index aligners (e.g., ERT-BWA-MEM2).
• Optimized for Modern CPUs: Features an Intra-query Parallel algorithm for the seed-extension phase, utilizing AVX2 instructions to eliminate computation bubbles caused by varying read lengths.

🔧 Technical Innovations

BWA-FastAlign revitalizes the traditional alignment pipeline with two core algorithmic contributions:

1. Multi-Stage Seeding (Hybrid Index)

   • Combines Kmer-Index, FMT-Index (Enhanced FM-Index with prefetching), and Direct-Index.
   • Dynamically switches strategies based on seed length and match density.
   • Achieves an 18.92× improvement in memory efficiency (bases processed per GB per second).

2. Intra-Query Parallel Seed-Extension

   • Unlike BWA-MEM2 (which uses inter-query parallelism and suffers from load imbalance), BWA-FastAlign parallelizes the Smith-Waterman alignment within a single query.
   • Includes Dynamic Pruning to skip zero-alignment scores.
   • Implements a Sliding Window mechanism to reduce costly memory gather operations.
   • Achieves 3.45× higher SIMD utilization, performing consistently well on both WGS (Whole Genome Sequencing) and WES (Whole Exome Sequencing) data.

📥 Installation

Prerequisites

• Linux operating system (tested on Ubuntu 22.04).
• GCC compiler (version 11.4 or higher recommended).
• CPU supporting AVX2 instructions (most modern Intel/AMD CPUs).
• zlib development files.

Compilation

git clone https://github.com/your-username/BWA-FastAlign.git
cd BWA-FastAlign
make

📖 Usage

BWA-FastAlign follows the same command-line interface as BWA-MEM.

0. Download Datasets. We download E.coli reference genome and sequencing reads.

# Download reference genome
wget http://ftp.ncbi.nlm.nih.gov/genomes/all/GCA/000/005/845/GCA_000005845.2_ASM584v2/GCA_000005845.2_ASM584v2_genomic.fna.gz
gzip -d GCA_000005845.2_ASM584v2_genomic.fna.gz
mv GCA_000005845.2_ASM584v2_genomic.fna ref.fasta

# Download sequencing reads
wget ftp://ftp.sra.ebi.ac.uk/vol1/fastq/SRR258/003/SRR2584863/SRR2584863_1.fastq.gz -O reads_1.fq.gz
wget ftp://ftp.sra.ebi.ac.uk/vol1/fastq/SRR258/003/SRR2584863/SRR2584863_2.fastq.gz -O reads_2.fq.gz

1. Index the Reference. Before alignment, you must index your reference genome.

# This will generate the hybrid index files
./fastalign index ref.fa

2. Align Reads (Mem). Map single-end or paired-end reads to the reference.

# Single-end alignment
./fastalign mem ref.fa reads.fq.gz > aln.sam

# Paired-end alignment
./fastalign mem ref.fa read1.fq.gz read2.fq.gz > aln.sam

# Using multiple threads (Recommended: 32-128 threads for high throughput)
./fastalign mem -t 64 ref.fa read1.fq.gz read2.fq.gz > aln.sam

3. Options. BWA-FastAlign supports the standard BWA-MEM options. Run ./fastalign mem to see the full list.

📜 Citation

If you find BWA-FastAlign is useful in your research, please cite our paper:

@inproceedings{fastalign2026,
  title={Faster and Cheaper: Pushing the Sequence Alignment Throughput with Commercial CPUs},
  author={Zhonghai Zhang, Yewen Li, Ke Meng, Chunming Zhang, Guangming Tan},
  booktitle={Proceedings of the 31st ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP '26)},
  year={2026}
}