HOW I ACCIDENTALLY CREATED THE FASTEST CSV PARSER EVER MADE

🧠 Key Highlights

• Origin Story: It started as a personal challenge to apply branchless programming to a real-world problem—CSV parsing, notorious for its unpredictable branching due to commas, quotes, and newlines.

• Naive vs Optimized: The initial scalar parser was clean but slow, plagued by branch mispredictions and cache misses. Sanix then pivoted to AVX-512 intrinsics, enabling parallel processing of 64 bytes at once.

• SIMD Magic:

  • Uses AVX-512 to compare characters in bulk.

  • Generates bitmasks to identify delimiters without branching.

  • Applies bit manipulation (_tzcnt_u64, popcntu64) to extract positions efficiently.

• Memory Mastery:

  • Employs memory-mapped files (mmap) for zero-copy access.

  • Advises the kernel to use huge pages for fewer TLB misses.

  • Aggressively prefetches data to keep caches fed.

• Node.js Integration: Wrapped the C parser using N-API, enabling blazing-fast CSV parsing in JavaScript with minimal overhead.

• Benchmark Results:

  • CLI: Beats tools like csvkit, rust-csv, and miller by wide margins.

  • Node.js: Outperforms papaparse and csv-parse by 2–4x in throughput.

• Real-World Impact: cisv can process 1TB of CSV data in ~10 minutes, compared to hours with traditional parsers.

• Caveats:

  • AVX-512 throttles CPU frequency and consumes more power.

  • Portability is limited to x86-64; ARM would require NEON rewrites.

• Philosophy: Optimization isn’t evil—it’s essential when you're in the “critical 3%” where performance truly matters.