Summary

  • The TMA 2026 study presented at RIPE 93 found ECN negotiation at 97.98% of reachable IPv4 web servers in its sample, yet 3.26% of ECN-aware IPv4 servers sent no expected ECT mark after negotiation.
  • More than 90% of measured paths preserved ECN bits. The minority that bleached them exposes a practical veto over congestion feedback; the study's seven AWS vantage points are a path sample, not an Internet census.

The Internet has learned a polite way to say that a queue is filling: mark the packet instead of simply throwing it away. RFC 3168 defines the ECN vocabulary. Yet a warning delivered with all the urgency of a blank postcard is still blank. That is the awkward finding behind Danesh Zeynali's RIPE 93 session.

In the original TMA 2026 paper, the researchers could negotiate ECN with 97.98% of the reachable IPv4 website servers in their sampled domain set; the equivalent figure for that IPv6 web-server set was 99.97%. Those numbers describe a successful conversation at connection setup. They do not say that a later packet carried a useful mark. The team checked: 3.26% of ECN-aware IPv4 web servers returned packets without the expected ECT(0) or ECT(1) marking after negotiation. For the IPv6 web-server sample, the failure was at most 0.5%. A handshake had said yes; the packet answered differently.

The route adds another clerk to the paperwork. Across the measured paths, more than 90% kept the ECN bits intact. In the minority that did not, bits were often reset to zero, and the observed bleaching clustered in relatively few autonomous systems. The paper reports 6.7% bleaching for its IPv4 paths and 3.9% for IPv6. Those are sampled paths from seven AWS vantage points and related probes, not a declaration about every packet or the intent of an entire operator. Even so, one eraser in a path can make an otherwise willing endpoint look deaf.

RFC 9330's L4S architecture asks more of congestion feedback: a useful low-latency signal has to survive the trip and reach a sender that responds. This study does not clock L4S performance. It identifies the unglamorous prerequisite on which that promise rests. The standards committee can approve the alphabet; servers, middleboxes and access networks decide whether the letter arrives.