NNSquad - Network Neutrality Squad
[ NNSquad ] Re: Routers deal with IP information
If we're talking about RED as the common mechanism used in routers on the Internet, see slide 33 of this research http://www.ittc.ku.edu/research/thesis/documents/mark_wurtzler.pdf. The graph shows a much lower drop probability for UDP versus TCP especially at the higher queue depths, though the drop probabilities are closer at lower queue depths. This would seem to back your original impression that TCP dropping was favored over UDP dropping it it's beyond the idea stage. This seems to make sense since the router traditionally can't assume that the UDP application will back-off as a response to packet drops. So even though uTorrent UDP does do a wonderful job backing off, this particular example of RED linked above which favors dropping TCP may be problematic until it's been reconfigured. And as you've pointed out, the reports from Canada have shown uTorrent UDP escaping traffic policing so we can't conclude it wouldn't be problematic elsewhere. Granted, it may just be a matter of reconfiguration to avoid these problems. The bottom line is that most of the stink about this debate could have been avoided if people stopped trying to tear each other's heads off. We can basically sum it up by saying. * No, BitTorrent was not "declaring war on VoIP, gamers" or DNS or other UDP applications * No, the concerns are not "utter nonsense" because the move to UDP can be problematic * Yes, uTorrent UDP does back off nicely but it does escape certain congestion control implementations * Yes, routers (which commonly implement RED) are configured to favor dropping TCP because of the traditional non-responsiveness of UDP applications to dropped packets. George -----Original Message----- From: nnsquad-bounces+george_ou=lanarchitect.net@nnsquad.org [mailto:nnsquad-bounces+george_ou=lanarchitect.net@nnsquad.org] On Behalf Of Richard Bennett Sent: Tuesday, December 02, 2008 7:46 PM To: David P. Reed Cc: nnsquad@nnsquad.org Subject: [ NNSquad ] Re: Routers deal with IP information Unfortunately, I have to agree with David Reed on this point. I was formerly under the mistaken impression that IS-IS routers preferred to drop largish TCP segments over UDP packets and TCP ACKs until quite recently, but have learned that this thinking never advanced past the idea stage. Most of your IS-IS routers today simply implement RED and only apply different dropping strategies across streams with different CoS settings. In the Internet core, it's all pretty much the same. TCP and UDP may be treated differently by management systems such as the Ellacoya traffic shaper (apparently used by Bell Canada,) Sandvine, and others, but this is a feature that's under the owner's control. Reports from Canada indicate that uTorrent 1.9 does in fact escape traffic policing at the moment, much to the chagrin of P2P over TCP users and VoIP over UDP users. I expect this situation to change, but can't say in which direction it will go. RB David P. Reed wrote: > One more correction of a complete boner by "expert" George Ou: >> >> Because UDP end-points don't respond to dropped packets the way TCP >> end-points do, most routers leave the UDP traffic alone when there is >> congestion and they only drop TCP packets which respond by cutting their >> flow rate in half. > This is UTTERLY false. Perhaps if Mr. Ou worked for a router company > instead of making up fantasies in his own mind, he'd understand this. > (it may be that "NAT boxes" do such things, and are often called "home > routers", but routers do not - the generality of the statement is > awesome in its ignorance). >> Besides, there's no point in forcing a 30 kbps gaming >> UDP data stream to slow down because it's already very slow and it's >> only >> fair to ask the bursty applications that have no bandwidth limit >> operating >> at 100 to 500 times faster to take a hit on bandwidth. >> > Like most people who are not on solid ground, he has to give details > of "why" his fantasy must be true. >> Now here comes BitTorrent with their well-meaning but problematic >> change to >> take a bulk file transfer protocol and stick it on UDP. So instead >> of tiny >> 30-80 Kbps VoIP and online gaming UDP streams, we're now looking at >> multiple >> UDP streams operating at 15,000 Kbps per user. Now we're forcing the >> network operator to change their routers inside the Internet to start >> managing UDP flows by dropping UDP packets whenever a link is congested. >> >> > RTP streams such as streaming video are *designed* to handle dropped > packets. That's the one of the main reasons we invented UDP - in > particular, Danny Cohen, who created the field of packetized speech > over the Internet pretty much from the whole cloth, taught me why. > > Ou is close to impeaching his credibility here. -- Richard Bennett