[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[ih] Why did congestion happen at all? Re: why did CC happen at all?

On 02/09/2014 00:32, Noel Chiappa wrote:
>     > From: Brian E Carpenter <brian.e.carpenter at gmail.com>
>     > I vividly recall Ross Callon speaking about why QOS routing doesn't
>     > work ... using the analogy of dancing in your own shadow, with a
>     > practical demonstration that it can't be done.
> Err, no.

Well, nevertheless, he did, at the QOSR BOF at IETF 36 (Montreal).
Unfortunately there don't seem to be any slides on line.

> First, 'QoS routing' != 'load-dependent routing' (depending on how you define
> 'QoS routing'). 

Indeed. But that was a BOF, where the distinctions were being clarified.
Ross was mainly making the point about timescales (speed of light vs
speed of dancing feet).


> As John Day says, they are very different time scales. I take
> 'QoS routing' to mean 'different traffic groups have their paths selected
> independently, based on some characterics of the traffic _and_ links; and
> 'characteristics' can be an almost-infinite range, not just bandwidth and
> delay (which are so over-engineered these days that it's almost not worth
> thinking about them any more, in path selection). E.g. some traffic might
> want to keep off links which are in country 'X', whereas other traffic might
> be OK with going through 'X'.
> Although I suppose 'low delay' - the main measurable impact of congestion -
> might be seen as a QoS type, but to bring up John's point again, the
> time-scale can be different; there's a _big_ difference between _baseline_
> high delay (e.g. a satellite hop), which one would approach with classic QoS
> type mechanisms, versus _dynamic_ (i.e. queuing) delays, which are _so_
> dynamic that they almost inevitable will demand a specialized mechanism.
> Second, it is not at all impossible to do load-dependent path selection. The
> ARPANET showed it could be done (albeit on a 'smallish' network). TLi has
> alluded to sophisticated damping. But the best way to approach the
> instability issue is just to cut the Gordian knot and get rid of the feedback
> loop.
> One can do this if one uses a Map-Distribution/Explicit-Path-Selection
> architecture (one the many charms of that design quadrant). If one picks a
> path for flow X, and sets it up (doing reservations if needed), and _then
> doesn't vary it_ (unless a link fails), then by definition there cannot be
> any oscillation. As a nice side-effect, this also allows one to do 'pull' on
> the distribution of the load level information - in the setup - not 'push',
> thereby attacking another of the fundamental issues with large-scale
> load-dependent routing.
> This is a 'not enough rooom in this margin' gloss on a very complex issue
> (the interaction with the nice stat mux property that Vint alluded to is an
> issue, for example), but since this is the Internet-_History_ list, not the
> Network-_Design_ list, I'm going to stop there.
> Although of course, in practice (not theory), this is all irrelevant, since
> <sarcasm>as we all know, routing == BGP, and there is no such thing as
> non-BGP-based routing</sarcasm>.
> I fear the Internet has reached a stage of ossification, where nothing will
> change until something entirely new comes along, and does to it what the
> internal-combustion engine did to buggy whips.
> 	Noel