So I'm reading NANOG and finding a certain thread interesting in this context. Look in the archive (or if you follow it, just look at your archive) for the subject line "Spectrum (Charter) Fragmented UDP".
> On Oct 1, 2019, at 4:57 PM, Suresh Krishnan <[email protected]> wrote: > > Hi all, > The text changes in version -16 of this draft for the resolution of Alissa > Cooper’s DISCUSS position resulted in lively discussion on the mailing list. > The authors have made another revision of the draft (version -17) to address > the comments raised during this discussion. If you believe that any of your > substantive points have not been addressed please respond to this mail > stating what they are. As the responsible AD, I believe that the draft is > ready to move forward and be approved, and I would like to do so by end of > day on October 8th 2019 in the absence of any actionable objections. > > Thanks > Suresh The point made this morning is that telephones have horrible TCP implementations: > From: Phil Lavin <[email protected]> > Subject: RE: Spectrum (Charter) Fragmented UDP > Date: October 2, 2019 at 8:59:42 AM EDT > To: Saku Ytti <[email protected]> > Cc: "[email protected]" <[email protected]> > >> While we can say this should just work, the reality is, it's not very >> reliably true and I would not build product or business on the assumption >> that it works well. > > Yup. Understood. We can't get away from sending multi-packet messages. We try > our best to keep SIP messages as small as possible though sometimes certain > optional features required by customers push it beyond their MTU. We're also > starting to see decreasing MTUs as customers deploy various SD-WAN solutions > and it's tough to keep up with these when you're already teetering on the > edge of what used to be considered a fairly common minimum MTU value. > > We can, of course, get away from using UDP. We can and do run SIP over TCP > and indeed over TLS on TCP though the stateful nature of TCP often makes this > undesirable. We see a lot of SIP phone implementations that do not handle TCP > connection failures very well and result in a loss of calls for a period of > minutes if this happens, as they take a while to notice the connection has > dropped and should be re-established. Pros and cons of each. > > If anyone has any specific information about Spectrum CPE changes or indeed > any contacts who may be able to interrogate this internally within Spectrum > that would be appreciated. What we are saying in this draft is that such things ought not to be. Do we have practical advice in context? "Fix your TCP and use it"? _______________________________________________ Int-area mailing list [email protected] https://www.ietf.org/mailman/listinfo/int-area
