Ginevra Fabrizio: 11 Bytes Saved, NXDOMAIN Still Oversized

Summary

A one-day replay of .nl DNS queries gave unpadded Falcon-512 signatures a 655-byte median, 11 bytes below the fixed padded size. The smaller format did not meaningfully increase the number of NOERROR replies under the 1,232- and 1,500-byte limits.
Every tested NXDOMAIN reply exceeded both limits. Multiple signed NSEC3 proofs, rather than a single signature’s padding, made those negative answers the hard case.

The compressed signature won the byte-count contest and lost the one-packet contest. That is the sharp edge in Ginevra Fabrizio’s RIPE 93 talk on Falcon and DNSSEC: an 11-byte median reduction looked neat on a signature chart, then disappeared inside a much larger answer.

Fabrizio, a researcher at the University of Twente, presented a study built with SIDN colleagues. Its paper appeared at the Applied Networking Research Workshop in 2025; she brought the format comparison to the DNS session at RIPE 93 on 29 October 2026. The question was practical: when a DNSSEC signature is slightly shorter, do more real answers fit in one packet?

DNSSEC signs DNS data so resolvers can check a response’s authenticity and integrity. A negative answer may need several NSEC3 records and signatures to prove that a name is absent.

The team started with authoritative .nl query records captured over 24 hours on 3 April 2025. They removed the zone’s existing DNSSEC records, signed two test copies with Falcon-512, and replayed the recorded queries against NSD servers: one copy used padded signatures, the other unpadded. The study measured response sizes for the two common response classes, NOERROR and NXDOMAIN. It did not put Falcon into the live .nl production zone.

The signature numbers favour compression. Padded Falcon-512 has a fixed size of 666 bytes. Unpadded signatures in the test zone had a median of 655 bytes. Eleven bytes is real, but it is one part of a packet assembled from the query name, DNS records, DNSSEC proofs and their signatures.

For NOERROR answers, most responses stayed within the 1,232-byte DNS message budget or the 1,500-byte Ethernet MTU used for comparison. A smaller but non-negligible share crossed those limits. The padded and unpadded curves were close; the paper found no meaningful increase in answers that fit because the signature format changed. Most ordinary answers already fit, and an 11-byte median difference rarely changed which side of the line a response landed on.

NXDOMAIN exposed the less comfortable result. A signed negative answer has to prove that a name does not exist. In this zone, that proof could include two or three NSEC3 records, each with its own signature. The replay produced 113,348,878 NXDOMAIN responses, and every one exceeded both size thresholds. Replies with two signed NSEC3 records ranged from 2,269 to 2,420 bytes; those with three ran from 3,075 to 3,767 bytes. The signatures were passengers in a packet already carrying several proofs.

The NOERROR set contained 59,312,128 replayed NOERROR responses. Its threshold-crossing minority matters, but it should not be swapped with the NXDOMAIN result: the former was mixed, the latter cleared both limits every time in this test. The paper’s answer-size table makes the mechanism legible. Add another signed proof and the response jumps by hundreds of bytes; adjust one signature format and the change is much smaller.

Fabrizio’s paper found that variation in queried domain-name length was of the same order as the size variation of one unpadded Falcon-512 signature. That comparison is useful because it puts the compression gain back into DNS’s messy arithmetic. Names vary. Proof counts vary. A format that saves a few bytes on one signature does not control the shape of the complete answer.

The study’s conclusion favoured the fixed-size padded format for predictability and fewer implementation errors. In the RIPE 93 Q&A, Fabrizio described floating-point arithmetic as a difficult implementation concern and weighed it against substantial TCP fallback. Padding does not shrink an oversized answer; it keeps signature length predictable while the ecosystem weighs that complexity against a small average saving.

She also said she and SIDN were beginning a testbed for resolver-side and authoritative-side workloads using Falcon and ML-DSA. The results were too early for the talk, so the next useful evidence would be response-size distributions joined to actual resolver retry counts and latency. The paper did not measure live recursive retries or end-to-end delay; those facts could change the operational cost assigned to the 11-byte difference.

For now, the important split is plain: most tested NOERROR answers fit, while every tested NXDOMAIN answer remained oversized. Any future Falcon choice for DNSSEC should be judged by answer shape and proof count, not by the median signature alone.