On Saturday, November 07, 2015 03:55:36 PM Paul Hoffman wrote: > Greetings. draft-stenberg-httpbis-tcp covers TCP considerations for > HTTP. During the discussion of draft-ietf-dnsop-5966bis, we often said > "HTTP servers handle lots of TCP just fine", and this draft describes > how. Are there things in here that we don't cover that we should?
to the extent that tuning the number of file descriptors and tcp protocol control blocks is important, yes. to the extent that kernels who hash their tcpcb's vs. using simple linked lists will do better at demuxing, yes. to the extent that servers who use kevent will do better than those who use poll or select, yes. however, the basic persistence negotiation in modern http are the same as what 5966-bis covers, and so would be duplicative here. if we're seriously willing to reconsider any of that logic, i'd like us to address the connection ecosystem with phase changes. that is, a server whose connection pool fills beyond some threshold might first advise its client pool that the clients should avoid persistence, and give those clients a chance to gracefully shut down their connections as a way to indirectly manage the server's resources. only when the connection pool gets into its "red zone" should the server start initiating connection closes. i considered this to be a stretch for the current draft and its schedule, and so i wasn't going to bring it up. but if we're bringing up other stuff, we should talk about this too. > It's a > -00, and will probably not be published for real for quite some time, > but maybe an Informative reference to the draft in > draft-ietf-dnsop-5966bis would be useful to some readers. an informative reference sounds best. -- P. Vixie _______________________________________________ DNSOP mailing list [email protected] https://www.ietf.org/mailman/listinfo/dnsop
