[e2e] "congestion" avoidance...

Jon Crowcroft J.Crowcroft at cs.ucl.ac.uk
Tue Apr 17 02:06:58 PDT 2001



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.



More information about the end2end-interest mailing list