RIPE 93 ECN study finds 4.3% path bleaching

The Handshake Said ECN. The Network Ate the Ink.

Why it matters

A RIPE 93 study found that most sampled servers could agree to carry congestion signals. Then the researchers followed the packets: on more than two million successful traces, the signal bits were erased.

Daniel Kade27 October 2026 at 21:56 UTCReading Time 2 minLast update: 27 October 2026

Illustration: the handshake kept its word. The path had other ideas.

Summary

Researchers presenting at RIPE 93 observed ECN-related bits erased on 2,278,845 of 52,916,794 successful path traces: 4.3% of that measured population.
Bleaching appeared on 6.7% of successful IPv4 traces and 3.9% of IPv6 traces. An endpoint's successful ECN handshake answers a different question from whether the packet keeps its marking.

The server says it understands a congestion warning. The packet leaves with the right two bits. Somewhere down the road, someone rubs them out. That is the awkward middle act in Danesh Zeynali's RIPE 93 presentation and his team's 2026 measurement paper.

The paper reports 52,916,794 path traces that reached their targets across its web, university and IPv6 endpoint samples. In 2,278,845 of those successful traces, a router or middlebox cleared ECN-related bits. The 4.3% figure belongs to successful traces, not to all Internet packets. The split is sharper: 6.7% of successful IPv4 traces and 3.9% of successful IPv6 traces showed bleaching.

At the other end of the story, the researchers could negotiate ECN with 94.37% of sampled IPv4 destinations and 93.66% of sampled IPv6 destinations from all seven vantage points. Those are destination counts. They cannot be swapped into the path-trace denominator. The same paper found that 3.26% of its IPv4 web-server sample failed a basic ECN compliance check after negotiation. Three tests, three opportunities to declare victory too early.

The researchers' measurement site describes the path probes: four traceroutes per target, covering the ECN codepoints and an unmarked comparison. Seven AWS locations supplied the vantage points. Toward the sampled endpoint sets, a small group of networks accounted for much of the bleaching; in a separate random-prefix test, no single AS dominated. The source population changes the suspect list.

RFC 9330 builds the L4S low-latency architecture on ECN. A service pitch can point to a successful handshake. The running path still gets a vote. If it wipes the marking, the fancy congestion conversation has arrived with its microphone switched off.