[e2e] "congestion" avoidance...
Henning G. Schulzrinne
hgs at cs.columbia.edu
Wed Apr 18 09:56:51 PDT 2001
Jon Crowcroft wrote:
> i'm wondering whether this is because you have used conservative
> parameters for the access capacity?
There may be a different model assumption here. We assume sources that
request a given rate (either explicitly or implicitly), where the total
demand can be significantly greater than the capacity, but the total sum
of the per-stream acceptable bandwidths is smaller. The results seem to
indicate that demand matters very little, but there's obviously little
choice but to turn away streams, either by admission or poor service,
once the sum of the minimal demands exceeds capacity. You do tend to
turn away the streams that value the communication service the least.
> so my model is this
> we have a range of social models users can choose ferom - they can
> subscribe to a "social model" server - it could be like napster or
> better like gnutella, and offers a market in capacity for peoppe in
> that social - the SUM of the usersd capacity is "tcp like" but within
> this ,they get to choose,. based on which server they subscribe to
> one server might offer stable price over session life, anoher might
> offer dynamic price on a packet timescale.
> servers can implemnt a range of different strategies - e.g. it might
> schedule each persons session over a long period (i.e. n sessions arrive
> at once, but are scheduled to compelteion, one oafte th other serialy)
> or it might interleave them in burst,s or it might be RAP or TCP Like
> or otherwise....
> now, some users WANT to have different rules than others (e.g. a set
> of multiple sources in a game versus a set of independant web
> downloaders or browsers)
> so thisis a market of markets, where we make the current rule (TCP) a
> meta rule but peopel can have local divergence...
Makes sense. I think the primary problem with the usual
charging-vs.-flat, reservation-or-no-reservation debates (pick your
decade) is that there's an implicit assumption that all users are forced
to use a particular service, so one must defend one's preferred choice
to the end.
I would be happy to be given the choice of services, from
congestion-shared TCP to nailed-down-admission-controlled "circuits" for
a remote PhD defense. Currently, I have to choose different networks,
different terminals and different addresses to get these different
services. It's a pretty bad sign if even the IESG and IAB still use
circuit-switched technology for their teleconferences since they/we
(apparently) don't quite trust the Internet service model...
Henning Schulzrinne http://www.cs.columbia.edu/~hgs
More information about the end2end-interest