ECN Passed the Handshake. The Route Still Erased the Signal

Summary

Among successfully reached Tranco web servers, 97.98% of IPv4 and 99.97% of IPv6 endpoints negotiated ECN. That measures endpoint capability, not the whole route.
In a separate path sample, 2,278,845 of 52,916,794 complete traces showed ECN-bit bleaching: 4.3% overall, 6.7% for IPv4 and 3.9% for IPv6.

ECN can get a yes at the handshake and still lose its message before the sender can act on it. Danesh Zeynali’s RIPE 93 presentation makes that gap the subject of a measurement study: a successful TCP negotiation, an intact congestion mark on the route, and actual use of ECN in traffic are three separate tests.

The distinction matters because Explicit Congestion Notification is meant to give a router another option besides dropping a packet. An ECN-capable router can mark congestion in the IP header; the receiver reports the mark to the sender, which can reduce its sending rate. Low Latency, Low Loss, Scalable Throughput (L4S) also depends on ECN working across endpoints and network equipment. A handshake proves that two endpoints can negotiate the feature. It cannot certify every device between them.

The paper’s active endpoint tests used seven geographically distributed measurement points. For the Tranco web-server set, 97.98% of successfully reached IPv4 servers and 99.97% of successfully reached IPv6 servers negotiated ECN. The next check exposed a smaller break: after a successful negotiation, 3.26% of the tested IPv4 web servers failed the paper’s basic IP-layer ECN response test; the corresponding IPv6 Tranco figure was 0.06%. The endpoint sample is a measured web-server set, not a count of all Internet hosts.

Then the researchers tested what happened to ECN bits in transit. They sent traceroutes with ECN-related bits set and compared the values along the path. Of 13.2 million IPv4 traces and 64.4 million IPv6 traces collected for the endpoint lists, 8.4 million and 44.5 million respectively reached their destinations. Among those 52,916,794 completed traces, 2,278,845 recorded at least one bleaching observation. The rate was higher in IPv4 than IPv6. Resetting the ECN bits to 00 was the dominant change.

Those are trace counts, not unique users or a direct measure of application failures. Hop attribution also required adjacent responding hops around the change; a silent router leaves the exact point unresolved. The study’s path tests and endpoint handshakes are separate measurement sets, so the results do not describe one transaction followed from SYN to application response.

Passive measurements add a third view. In monthly MAWI snapshots from 2007 to July 2025, about 3.8% of TCP flows with successful handshakes included ECN negotiation from 2017 onward; ECN appeared more often in larger, data-carrying flows. In M-Lab tests spanning August 2021 to August 2025, mobile networks showed nearly twice the ECN negotiation rate by 2025 compared with 2021, but IP-layer ECT signals lagged—especially on mobile paths. IPv6 showed stronger ECT success across the measured ISP and mobile networks.

The practical lesson is a measurement one: “ECN enabled” is too blunt a status field. Zeynali’s team separates capability at the server, preservation across the path and use in observed traffic. Each answers a different question, and only the last two reveal whether the congestion signal can survive beyond the handshake. The paper measures support and signaling; it does not measure end-user latency or throughput gains.