Multipath TCP for Linux (2022)

Original Article ↗ Hacker News Discussion ↗

Adoption and Kernel Design

  • Linux requires explicit application opt‑in; maintainers rejected a fully transparent default.
  • Some argue transparent MPTCP (automatic subflows, e.g., LTE failover) would cover most use cases.
  • Servers and even whole GNU/Linux distros could enable MPTCP by default via eBPF, but this is not widespread.
  • Several commenters express disappointment at slow adoption despite early promise (since ~2013).

Middleboxes, Compatibility, and Security

  • Most proxies/firewalls reportedly pass MPTCP options; main hurdle has been load balancers.
  • MPTCP falls back safely if options are stripped, assuming middleboxes don’t enforce contiguous TCP sequence space.
  • Concerns raised about new security issues when multiple IPs are used under one logical connection (e.g., IP-based ACLs).

Use Cases and Real-World Deployments

  • Strong interest in seamless Wi‑Fi/cellular handover and resilience (e.g., “parking lot” Wi‑Fi failures).
  • Some mobile apps (e.g., Siri, WeChat) and iOS support MPTCP; one user notes turning it off due to metered cellular data.
  • Delivery robots and multi‑modem setups seek <50ms failover; commercial solutions (Peplink, Phantom Auto) are used due to engineering complexity.
  • European ISPs reportedly use MPTCP proxies to combine weak DSL with cellular for rural customers.

Routers, Home Networks, and Aggregation

  • OpenWrt derivatives (e.g., OpenMPTCProuter) and Home Assistant OS now support MPTCP.
  • Users report aggregating multiple home links (cable + fiber, dual 5G) via a VPS, but note drawbacks (e.g., datacenter IP at home, TCP tunneling HoL issues).
  • Some argue home HA and aggregation are already solved with LACP/VRRP; others value per‑flow bandwidth aggregation and smoother roaming.

Relation to SCTP, QUIC, and MPQUIC

  • SCTP is cited as an earlier, similar idea with limited public internet traction but heavy telco use (and in WebRTC data channels).
  • Several see innovation shifting to QUIC because TCP extensions are often broken by middleboxes.
  • Others view QUIC as a step backward (UDP issues, initially no multipath); a standards‑track multipath QUIC draft and implementations exist.
  • MPQUIC aims to combine MPTCP‑style multipath with QUIC’s stream multiplexing.

Architectural Reflections

  • Some wish TCP had connection IDs independent of IP addresses to allow native roaming/multihoming.
  • Debate over whether such features belong at transport vs. higher layers; references to Mosh and session‑layer concepts.
  • Overall sense: MPTCP is technically valuable but hampered by ecosystem inertia, middleboxes, and load‑balancing complexity.