Bypassing the kernel for 56ns cross-language IPC

https://github.com/riyaneel/Tachyon/tree/main/docs/adr

Article

  • Tachyon achieves 56.5ns p50 round-trip IPC between C++, Rust, Python, Go, Java, Node.js.
  • Uses shared-memory ring buffers with futex-based signaling, bypassing kernel on the hot path.
  • No serialization overhead; targets RAM-speed communication between independent processes.
  • Beats Aeron in benchmarks for small payloads (32 bytes).

Discussion

  • Author present in thread; confirms design goals and benchmarks.
  • Purists note futex use means the kernel isn’t fully bypassed; it’s hot-path only.
  • TOCTOU attack surface flagged: sender can manipulate shared memory after receiver reads header.
  • Go binding uses cgo, which disappointed some; pure-Go or purego alternative requested.

Discuss on HN


Type Link
Added Apr 20, 2026
Modified Apr 20, 2026