[e2e] "congestion" avoidance...

Jon Crowcroft J.Crowcroft at cs.ucl.ac.uk
Tue Apr 17 04:38:21 PDT 2001

In message < at mail.reed.com>, "David P. Reed" t

 >>I think I was trying to "napsterize".
yes, i agree with this, but i was going further - that we don't need
to have global per source/dest agreement - it can be on arbitrary
aggregations (a la RSVP wildcard matching type stuff) and then withi
nthose aggregations, the tradeoffs could be agreed differently - 
 >>The seed of the idea is that you get all kinds of signals as a result of 
 >>congestion that can be used in a control loop, so why not add a simple 
 >>signal that can be modulated on the basis of source/dest intent, and which 
 >>is shared with other participants in the congestion.  (It's like money in 
 >>the sense that its scarcity means you only want to use it at a rate you can 
 >>Besides your packet rate (which directly affects other flows in the 
 >>"flux"), having a bit per packet input that is combined into the other 
 >>flows' flags would seem to be a powerful differential signalling channel.
 >>Let's say that a normal flow has about 50% of its packets with "urgency" 
 >>on, 50% off.
 >>Say when you turn on ECN in a packet with the "urgency" flag set, you turn 
 >>off the next "urgency" flag you see on a packet arriving into the 
 >>queue.  Then the source/dest of the other flow sees that it is interfering 
 >>with an urgent flow because it sees that its own urgency rate has 
 >>dropped.  [This is the sketch of how a cross-flow signalling scheme might 
 >>provide information from one flow to others - the algorithm is not intended 
 >>to work, of course.]
 >>At 11:20 AM 4/17/01 +0100, Jon Crowcroft wrote:
 >>>so the problem i have with the current "ECN pricing" thinking is that
 >>>it ignores users preferences for stability and predictability over
 >>>cheapness (and we have a LOT of evidence gathered from mobile phone
 >>>contracts, web and traditional telewphony as well as airline procing
 >>>and so on i can cite)
 >>>the shadow price for a packet (smart market) is one model, but leads
 >>>to potential rapid fluctation in price aroudn flash crowd periods,
 >>>which are all to common in IP networks
 >>>the simple alternative  is a shadow price for a virtual circuit - this
 >>>gives a stable price for the throughput over the "lifetime" of a
 >>>if you Mix this, you can get into nasty arbritrage situations....and i
 >>>don't buy into the story that yo ucan offer users a choce via risk
 >>>brokers - for the very reason that the traffic is highly dynamic...
 >>>also, risk brokers form markets  themselves.....
 >>>what i was thinking ewas to "democratise" (disintermediate) the risk
 >>>broker and let users form their +own+ cartels dynamically...
 >>>i.e. we napsterise congestion pricing for packets and flows...
 >>>In message < at mail.reed.com>, "David P. 
 >>>Reed" t
 >>>  >>Interesting thoughts.  However, money or something like it needs to enter
 >>>  >>into the thinking.  I.e. some notion of sharing responsibility for costs
 >>>  >>imposed on others.
 >>>  >>
 >>>  >>IE: At a point of congestion, the "indirect channels" among competing 
 >>> flows
 >>>  >>provide a way of signalling (at some bitrate) for a bargaining scheme.
 >>>  >>
 >>>  >>What range of bargaining schemes can be piggybacked on this signalling 
 >>> channel?
 >>>  >>
 >>>  >>For example, what if a single (urgency) bit per packet (like the ECN 
 >>> flag,
 >>>  >>but provided by the source to the congested queue) could be modulated at
 >>>  >>the source, tracked in a state variable at a router queue, and coupled 
 >>> into
 >>>  >>a bit in each outgoing packet that controls rate like ECN.
 >>>  >>
 >>>  >>
 >>>  >>At 10:06 AM 4/17/01 +0100, Jon Crowcroft wrote:
 >>>  >>
 >>>  >>
 >>>  >>
 >>>  >>>there have been some steps recently to look at a range of rate and
 >>>  >>>window based mechanisms for sharing the net amongst a set of sources (or
 >>>  >>>sinks if we include receiver based multicast schemes) - i was looking
 >>>  >>>at these and wondering if it isnt time to revisit some of the
 >>>  >>>congestion control and avoidance thinking
 >>>  >>>
 >>>  >>>some schemes have been proposed that smooth the adjustment so that
 >>>  >>>over an RTT we creep up to the operating rate, and creep down, on a
 >>>  >>>packet by packet (inter-packet delay adjustment) basis
 >>>  >>>(RAP from Handley et al)
 >>>  >>>
 >>>  >>>other schemes have proposed different powers for the increase decrease
 >>>  >>>function (and assert that so long as we decrease x^n, and increase
 >>>  >>>x^(n-1), we ought to be "ok" for some definition of ok)
 >>>  >>>(binomial adjustment etc from
 >>>  >>>
 >>>  >>>the TCP AIMD with fast retransmit scheme has several motivating factors
 >>>  >>>some intended, some lucky happenstance (serendipitous)...
 >>>  >>>
 >>>  >>>1/ sampling network conditions and eliminating noise:
 >>>  >>>
 >>>  >>>currently, this operates over the RTT timescale, but is memoryless
 >>>  >>>after that....estimates for loss effectively im,plicit in the AIMD
 >>>  >>>operation, but the noise filter (number of dupacks) is somewhat
 >>>  >>>rigid...
 >>>  >>>
 >>>  >>>2/ safe/stable operation:
 >>>  >>>given feedback controller, its reasonable to operate this over
 >>>  >>>packet conservation/self clocking makes it more smooth
 >>>  >>>
 >>>  >>>3/ relating end system rate adjustment timescale to buffering 
 >>> provisioning
 >>>  >>>the AIMD scheme has the bandwidth/delay product wrth of network
 >>>  >>>buffering as a necessary side effect - other adjustment schemes might
 >>>  >>>need less (some might need more but that almost certainly means they
 >>>  >>>are trouble:-)
 >>>  >>>
 >>>  >>>4/ social coupling - we have a target operating point which will be
 >>>  >>>some fraction of a bottleneck link
 >>>  >>>if we take a flow f, and a flux (sum of flows into a bottleneck) F,
 >>>  >>>then the idea is that we get a share proportional to the _resource_ we
 >>>  >>>use, which (approximately) includes 1/RTT as a factor (kelly et al, le
 >>>  >>>boudec et al)  the idea is that a set of fs in an F are coupled by the
 >>>  >>>loss or ECN feedback function, and by some reaction period being at
 >>>  >>>least in the same ballpark....
 >>>  >>>
 >>>  >>>in fact, though we don't have to have smooth functions at all, nor do
 >>>  >>>we have to sample only the average loss rate, nor choose the sample
 >>>  >>>rate to be an RTT - the RTT is a way of _loosely _ coupling things,
 >>>  >>>but is perhaps too strong
 >>>  >>>
 >>>  >>>what if someone wanted a _rate_ that persited for all (or a larger
 >>>  >>>part) of a connection? how could we work out some sort of congestion
 >>>  >>>model that accommodated both packet and connection timescales?
 >>>  >>>
 >>>  >>>at least one factor seems missing, and that is some estimate of the
 >>>  >>>number (and rate of change of number) of flows....if we alter the
 >>>  >>>sample period, and sample bot hte hcongestion feedback Mean, AND its
 >>>  >>>variance, we might be able to (assuming the social coupling function
 >>>  >>>was still "social") estimate this
 >>>  >>>
 >>>  >>>obviosuly if people want to they can behave anti-socially (but that is
 >>>  >>>and wil lalways be true unti lwe do pricing or othewr forms of
 >>>  >>>admission control) - letsassume they behave "nicely"....
 >>>  >>>
 >>>  >>>could someone choose to operate a "very slow" congestion control
 >>>  >>>scheme? why not? lets say i run a connection that takes 1/10 of the
 >>>  >>>capacity, but there are 5 other connections, then why should i react
 >>>  >>>to loss unless my longer term loss (or ecn) rate  tells me that
 >>>  >>>there's now 9+ other flows? currently,  if i run any adjustment
 >>>  >>>scheme based just on average, i have a chance of adjusting wrongly...
 >>>  >>>
 >>>  >>>more importantly, maybe
 >>>  >>>secondly, how about re-examining the social coupling function? why
 >>>  >>>shouldn't ten people _agree_ a different congestion partition function
 >>>  >>>(e.g. they have an application that requires n sources)
 >>>  >>>
 >>>  >>>i guess this could be implemented via the Congestion Manager type API,
 >>>  >>>but i am interested in the general family of functions that fit this
 >>>  >>>more general model - for example, it seems to me that you can have
 >>>  >>>radically different increase/decrease if you have
 >>>  >>>a) a different sample period and a more accurate deascripotion of the
 >>>  >>>evolutuon of the loss/load process over time (e.g. some sort of fancy
 >>>  >>>bayesian predictor)
 >>>  >>>b) a different share/social function - e.g. if we have 10 sources
 >>>  >>>agreeing on a different load, then how do they distribute this
 >>>  >>>information and how do  we make sure they aren't penalized by any
 >>>  >>>extra fancy stuff people might later add!
 >>>  >>>
 >>>  >>>j.
 >>>  >>
 >>>  cheers
 >>>    jon
 >>- David
 >>WWW Page: http://www.reed.com/dpr.html



More information about the end2end-interest mailing list