Vaishnavi Raghavajosyula: 76,700 Paths, One Edge Bottleneck

Summary

Vaishnavi Raghavajosyula and her co-authors measured 76,700 paths with RIPE Atlas. A 1,500-byte PMTU appeared on 80.5% of the sample; smaller limits were concentrated near network edges.
The study found 97 distinct values and an 18.5% mismatch between measured PMTU and MSS-inferred PMTU. A single comfortable default can waste capacity; a larger packet can still fail where path feedback is lost.

Packet sizing is a quiet tax. Choose too small and every transfer carries more packet overhead than the path requires. Choose too large and a packet that cannot pass may disappear without the warning the sender needs.

Vaishnavi Raghavajosyula’s RIPE 93 presentation put numbers on that trade. Her team measured 76,700 paths spanning access networks, the Internet core and web-serving infrastructure. The method used a traceroute-based search that did not depend on ordinary Path MTU Discovery or TCP’s Maximum Segment Size. That matters because the study was trying to see where packet size breaks, including paths where the usual feedback is missing.

The largest bucket was familiar: 1,500 bytes appeared on 80.5% of paths. Another 11.7% measured 1,492 bytes, and 1.5% measured 1,480. Across the sample, the researchers saw 97 distinct PMTU values. The pattern was not evenly spread. Sub-1,500-byte limits clustered near network edges; core paths rarely carried the same constraint. In access paths heading toward the core, the defining break often appeared within the first or second hop. Among core-path breaks, 75% occurred in the AS before the destination AS.

The route also has two directions. Raghavajosyula said reverse PMTUs were higher in most measured pairs. The paper abstract adds a practical warning: measured PMTU and a value inferred from TCP MSS disagreed on 18.5% of paths. A connection’s advertised segment size is not a clean substitute for the packet ceiling it will meet along the route.

The result is a map, not a universal setting. It supports more informed, path-specific probing and gives protocol engineers a better starting distribution than a slogan about the smallest safe packet. It does not show how many users would see faster downloads, lower latency or fewer losses under a new default. That would require a traffic-weighted production trial.