[Iccrg] Draft minutes
michawe@ifi.uio.no (Michael Welzl) Wed, 01 August 2012 20:26 UTC
From: michawe@ifi.uio.no
Date: Wed, 01 Aug 2012 20:26:21 +0000
Subject: [Iccrg] Draft minutes
Message-ID: <D67D3E1A-548A-45DE-AF8B-C5B38EC8DB17@ifi.uio.no>
X-Date: Wed Aug 1 20:26:21 2012
Hi all, Please find below the minutes from the ICCRG meeting last Monday, here at the ongoing IETF. Speakers, everyone else who was there, I'd appreciate if you'd take a look and send me any fixes you think are necessary ASAP. I think you can actually edit them directly here: http://etherpad.tools.ietf.org:9001/p/notes-ietf-84-iccrg?useMonospaceFont=true&showChat=false I'll make end of this week the deadline, i.e. I'll finalize the minutes next Monday, 6 August. Cheers, Michael **************************** Yuchung Cheng Client-aided Congestion Management for TCP Yuchung's presentation and slides described motivations, the overall idea of doing CC across connections through CM, and pros/cons of sender-side CM. He described some reasons why the client may be in a good position to help, which is what the Great Snipe client-based CM framework is about, as the subsequent slides describe. There are a number of research issues and opportunities. Great Snipe is in early stages; source code is planned to be released as it matures. [Jim Gettys] - not one, but 2 bottlenecks in the home, 802.11 and broadband. inside Linux and other OSes, there is a system that tracks how fast a wireless link can go (minstrel), on a 50ms by 50ms basis, you have some idea how fast you can transmit; thinking about this as a black box without that information doesn't make sense to me [Yuchung] - all this information is on the client-side or end device and if we can leverage it to help the transport layer that would be great; it is very difficult to do if we keep it inside one connection [Roland Bless] - didn't quite get how this differs from RFC 3124 (congestion manager). [Yuchung] - was imagining that was based on server-side; the problem is to share between servers within a farm is hard, so the client may be a better point [Tim Shepard] - had trouble following proposal; sender/server and client/receiver terminology [Yuchung] - trying to make a distinction NOW, in my talk, the direction of data flow as always server to client [Tim] - wasn't sure you've included upload case; I have a server at home; wasn't sure this would work in general, or what you intend [Yuchung] - just picking up the common case of a web server sitting in the cloud and a client behind a wireless or other link [Scott Bradner] - CM RFC is not limited to servers, but it is sender-side. As I recall, the reason CM didn't take off is that there was no way to make sure the same paths were being taken for each of the flows being managed [Matt Mathis] - this is early formative stage; it's probably the case that you need to do CC from both ends, which does (I believe) fix some of the CM problems. For many use-cases you potentially control all applications on a single bottleneck; it's an improvement even if you do better some of the time and doing classical CC at the other end fixes the problems [Michael Welzl] - what are you planning to put into code? [Yuchung] - still experimenting; quick answer is that I don't know yet [Matt] - still some major holes in the code Grenville Armitage FreBSD-based TCP CC research: Delay-Gradient TCP Modular TCP Framework Presentation and slides describe work done on the FreeBSD TCP stack, and aim to look for collaborators, testers, etc. The idea of delay-based TCP sensing onset of congestion through variations in RTT was described, in the "CDG" algorithm (CAIA Delay Gradient). Many aspects of the design and some results comparing to other algorithms were shown. Code is available in FreeBSD. [Question] - impact of RED [Kathie Nichols] - since you're looking at how RTT goes up and down are you inverting fairness between different RTTs when you have flows with different RTTs. [Grenville] - haven't looked at it [Kathie] - undershooting and getting worse throughput is another thing to look at [Grenville] - CDG will never obtain the same goodput that loss-based algorithms will; we will always underestimate a bit. the benefit is that you can get pretty good throughput but will hammer the voip and other interactive traffic less [Kathie] - comparing against reno instead of cubic; cubic is pretty widely used [Grenville] - we don't have it in these slides, but showed that cubic will hammer the link and is not a real win for home links [Andrew McGregor/Ken Carlberg] - packet aggregation would lead to fluxuation in RTT [Mo Zanaty] - you want to operate under the no-loss case and be parameterless; how do you reconcide those things? in what cases are you able to stay fully lossless [Grenville] - there are a couple of magic numbers; we don't actually revert to newreno per-se, we just use our shadow window to only back off as much as newreno would have under the circumstances [unknown 1] - how do you distinguish between congestion and non-congestion loss [Grenville] - a congestion loss happens when the delay gradients tell us that there was congestion before the loss. there's a potential weakness in incorrectly estimating that queues are still growing [unknown 1] that may not be the wrong answer; on UDP RTCWEB side people think a few percent loss is fine without reducing the rate; not sure anybody has really analyzed this; likely related to bursts of other flows starting up [Matt Mathis] - have you looked at worst-case scenario (10s of thousands of streaming customers) and a near-field bottleneck where loss is correlated across flows [Grenville] - haven't specifically looked; we would see a massive gradient [Matt] - jumping up assumes that I saw an old RTT; suppose I have a new connection [Grenville] - would have to think about that [Andrew/Ingemar] - non-congestive loss in wifi and LTE is about 40% in the raw channel, and it has its own mechanism for retransmission to hide its underlying packet losses [Grenville] - those turn into latency [Andrew] - the scarce resource is permission to transmit at all, so it does lead to latency but not as much as you think [Bob Briscoe] - cubic hasn't got enough control signals at speed agreement Microphone switched to Lawrence Stewart to discuss the modular CC framework in FreeBSD. Discussed FreeBSD as an R&D platform. The implementation of the modular CC capability was discussed, inclucing SIFTR, khelp/hhook [Andrew/jabber-Michael Tuexen] - does it include SCTP? [Lawrence] - TCP immediately is the focus, other protocols down the track [Andrew/Dave Taht] - when will we see codel in BSD? [Lawrence] - soon; several people working on it Matt Mathis Laminar TCP and Related Problems Matt discussed the current status of linux patch for Laminar, and the list of possible add-ons. Since some folks were in neither TCPM nor Matt's presentations at the last IETF meeting, he gave a brief overview of Laminar that isn't in the slides. [Michael Welzl] - what is in the patch; full-featured TCP? [Matt] - yes, it's essentially identical, except for a case where it's slightly slower, related to congestion window validation, and it doesn't do TSO the same way Matt encourages people to play with the code and will find that Laminar makes things better/easier. Wants to hear how it effects other people's work (making it easier or harder). laminar@ietf.org mailing list [Lars Eggert] - we have time to talk about what we want to do here that we didn't in TCPM [Mirja] - is this going upstream? [Matt] - this breaks anyone who knows the code well's intuition about how the code works; breaking it into smaller pieces now, but that intermediate kernel would require different debugging, so it's more work to do it that way. on the other hand, I have a history of long-term support of patches off to the side that eventually got imported Jim Gettys Jim discussed bufferbloat global research topics involving measurements needed to tell if it's getting better/worse and where it is. Netflix streaming behavior is to pump out chunks of data every 10 seconds, not actually streaming wants an adaptive AQM; Jim says he's throwing out questions, but he doesn't have answers. This is research to be done. work on tools is needed [Lars Eggert] - go to homenet and write a document about what needs to be done [Andrew McGregor] [Dave Taht] - openwrt many models already supported [Lars] - I need one that does fiber [Jim] Wireless is often as bad as broadband or worse Meeting adjourned 3PM local time
- [Iccrg] Draft minutes Michael Welzl