> Re: protocol changes since the -01 version of the base draft, MTL > no longer uses EDNS(0) to request full versus condensed signatures > [1]. Every RRSIG in the base draft is now a full signature containing > both the signed ladder (ML-DSA-44 signature) and Merkle proof. > Beyond what any new DNSSEC algorithm requires (code points, validator > support), no DNS protocol changes are needed. This simplifies the > base implementation with the trade-off of moving potential bandwidth > savings to an optional extension draft [2] while still giving > operators benefits for space and compute. > > Re: implementation complexity MTL does add some complexity compared > to ML-DSA-44, but the latest base draft is meant to ease that. For > validators, MTL is deployed like any existing DNSSEC algorithm > where each RRSIG can be evaluated against a DNSKEY. This change > has removed the complexity of needing to keep track of other > signature state as a validator has access to the full signature in > each RRSIG. Validators may still want to add logic to store the > same signed ladder only once to avoid re-verifying a previously > seen signed ladder in a full signature, but this is not required. > Another small bit of complexity is the need to implement Merkle > proof validation logic, but that would only add a few fast hash > operations.
I was curious, it is possible to only implement the base draft. Not from a protocol point of view, there they are independent, but operationally. What is the effect of only implementing the base draft on performance and how does that compare to ML-DSA-44. The first thing to note is that traffic patterns will be similar for MTA and ML-DSA-44 (number of TCP connections required) but the amount of data transferred will be somewhat larger. For a resolver that just implements MTA as a DNSSEC algorithm without further optimizations, memory use will be higher. Validation will take slightly longer. All in all, this is just negative compared to ML-DSA-44. On the authoritative side, if we take a traditional signer and an (anycast) distribution network. For a completely naive signer, the overhead would be higher (every signature would be a separate merkle tree). We can assume that the signer just has to be modified. In that case, memory requirements for the signer can be smaller and cpu time required for signing is less (because there will be only one PQC public key signature for a collection of RRsets). As far as I know, the rate of PQC signatures would not an issue in this setup if ML-DSA-44 is used, so this is not a real benefit (I could be wrong, does anyone have data that shows otherwise?) Which leaves memory reduction as main benefit. If we assume secondaries that only implement the base protocol without optimization then memory and bandwidth requirements will be higher compared to ML-DSA-44. With optimizations, memory use can be reduced but bandwidth requirements remain higher (assuming resolvers only implement the base protocol). This is a benefit for operators where memory is a bigger issue than bandwidth. For individual zones, there are not that many zones that don't fit comfortably in the memory of a mondern server. My conclusion is that though at the protocol level, the base draft is independent, operationally we need to implement optimizations that don't fit into the traditional DNS model. For example, recursive resolver that does not validate may still need to optimize caching of condensed MTL signatures or even request condensed signatures to reduce network bandwidth. So we need to carefully define what benefits we expect to get from using MTL, verify that in pratice we will get those benefits and that those benefits warrant the extra complextiy. I'm quite worried how much harder it will be to create future DNSSEC implementations. So far signature generation and validation was relatively simple and could be left to existing crypto libraries. MTL changes that, applications will have to create complex datastructs to store condensed signatures. One angle that requires attention is the potential for attackers to massively change traffic patterns. On it's own, MTL is not much worse than ML-DSA-44. However, if we expect lower memory and lower bandwidth from condensed signatures and an attacker can force a resolver to request and store full signatures then this may cause the operational advantages to be lost. _______________________________________________ DNSOP mailing list -- [email protected] To unsubscribe send an email to [email protected]
