Henning wrote > As we learned from SCTP, getting the first implementations of a > conceptually new protocol right is non-trivial. The goal would be to > have somebody build a library, so that most application writers don't
> have to worry about getting RTO estimation right, dealing with window > adjustments and all the other details that any non-toy transport > protocol needs. Thus, generality is an advantage, as it encourages > library building. Having done a user space implementation of TCP some years ago, I'd say that it is easy to get it running, but hard to get a really good TCP with decent throughput. Re-transmission tends to hide your bugs. Modern TCP implementations use a number of optimizations that are not documented in RfCs. As a user-space TCP needs timers, the networking code will be more complicated than ordinary socket code. Many OSes started with a BSD TCP stack and ran into unexpected difficulties when they tried to replace it with a home-grown solution. TCP-over-UDP is not exactly a lightweight solution.. Wolfgang _______________________________________________ P2PSIP mailing list [email protected] https://www.ietf.org/mailman/listinfo/p2psip
