[e2e] Skype and congestion collapse.

Jon Crowcroft Jon.Crowcroft at cl.cam.ac.uk
Sat Mar 5 08:05:19 PST 2005


one of the big differences between audio and video 
is that people can leave video running and often pointing out the window at 
"interesting" scenary (e.g. a road:)
but with audio, you tend to want to engage in a conversation with someone at the other end, and if the quality aint
good enough, you get voluntary congestion avoidance - users "hang up"
(as 
"Alexandre L. Grojsgold" <algold at rnp.br>
just pointed out)

so the point is that with lots of voice calls, the aggregate set of flows
will end up looking something like a big fat TCP,  varying slowly a people join and leave (when they've said their
piece)

of course, one nice thing with the mbone/multicast was that if the _receivers_ all left, the 
traffic stopped flowing ,so if someone was just plain boring, then you could stop receiving their traffic and
congestioning your inbound link - this is something that some of the anti-multicat lobby "failed to get"
about the value of the multicast model....someone could send, but if there were no receivers, the traffic didnt
go beyond the first router (at least once we got past the earlier multicast flood/prune)

anyhow, the statistcs in countries with very cheap (or free local)
 phone calls are people spend around 300 minutes a month talking...
so you can workout that thats some kind of social limit to the traffic

naively if you spend more than 50% of your waking life asleep^H^H^H on the phone, then you start to have no life
to speak of:)
(mind you that doesnt stop some people spending more than 100% of their  lives sending email...oh well,
meaning is such a scarce commodity anyhow...unlike hydrogen or stupidity (one for lloyd wood to spot there)).

cheers
j.

btw, last time this came up for me, I accused an ISP that blocked UDP of being _underprovisions_
and they set the lawyers on me -this would be back >10 years ago - nowadays we know what provisioning IP means, but
back then I was right  - provisioning was something phone companies did for voice, and
saying someone was underprovisioned meand "underprovisioned for voice" - most phone companies are overprovisioned
for voice, just like most ISPs now are overprovisioned for TCP/web (though perhaps not P2P) - for me, this means
that they have enough that an average interactive web session does not exceed around 1 second per web page - thats 
a 1970s result on human interaction with form entry systems - a modest
web page is somewhere around 12kbytes these days, so you are looking at 72kbps

so if they can't hack 10kbps skype, they are pretty much out of order w.r.t just about any useful internet usage
whether its "TCP friendly" or "selfish" or "inelastic"

btw, TCP friendliness _is_ selfish - its in your own interest to do as you would be done by.
and inelastic is not necessarily selfish - just constant sending even when the packets are not gettng through that
is stupid...

thought:
one solution to DOS would be for us to get rid of unicast and only use multicast groups of 1, then a dos attack
would be immediately removed by leaving the group:) 
(oh, i know, the attacks would just move onto the PIM RPs or onto the address allocation servers yeah...)


In missive <20050304205159.GF5440 at sbrim-w2k02>, Scott W Brim typed:

 >>On Fri, Mar 04, 2005 04:22:36PM -0300, Alexandre L. Grojsgold allegedly wrote:
 >>> Well, one must have in mind that at the end of a skype voice connection 
 >>> ther is as human being that, in case of congestion and excessive packet 
 >>> loss will simply hang up the call - or even be so unhappy that he will 
 >>> never try calling thru Skype again.
 >>
 >>When I was playing with low bandwidth video we backed off the transmit
 >>rate only so far -- below a certain point it wasn't worthwhile, so we
 >>stayed at our minimum and accepted whatever happened.

 cheers

   jon



More information about the end2end-interest mailing list