[Iccrg] Heresy recapped
lars.eggert at nokia.com
Thu Apr 17 22:13:07 BST 2008
On 2008-4-16, at 12:58, ext S. Keshav wrote:
> On Apr 16, 2008, at 1:03 PM, Lars Eggert wrote:
>> I wouldn't go quite so far and call the current status quo of TCP-
>> friendliness dogmatic; it's mostly what VJ congestion control left
>> us with, and it's sort of a better-than-nothing style of fairness.
>> I do encourage folks to revisit that design choice and discuss
>> whether this type of fairness is what the current and especially
>> future Internet should strive to adhere to, or whether there are
>> alternatives or generalizations that offer more flexibility or
>> "stronger" notions of fairness.
> I think two points are obvious:
> 1. VJ congestion control, or any end-point congestion control for
> that matter, cannot by itself provide fairness of any sort. Fairness
> (whatever the definition) must come from inside the network, simply
> because the edges are uncontrolled. We can argue all we want what
> fairness means, but this fact remains. So the status quo is broken,
> a fact that has been known for 20 years.
I completely agree, *if* the end systems cheat, and didn't mean to
imply that a purely edge based control system can deal with this.
But in many cases the edges don't cheat and VJ congestion control does
leave us with a situation where flows end up sharing a path. They may
not end up sharing it according to a number of stringent definitions
of fairness, but typically, all flows make progress. That's better
(I'm open to the argument that it gets easier and/or more attractive
for the edges to cheat than it used to be, and that that's why we need
a network-based mechanism..)
> 2. Even if we put into place mechanisms for 'fairness', it is not
> clear that anyone cares about fairness in the first place. Two
> endpoint owners can compare notes about what service they receive,
> but the reason why they get different service could be
> misconfiguration, different RTTs, bugs in their software, line noise
> (or RF noise), cross traffic or just the phase of the moon. As far
> as I can observe since 1988, the need for fairness is always
> assumed, rarely verified, and never necessary.
I think your definition of fairness is a lot more stringent than mine.
> My conclusion: fairness can't be achieved by endpoint control and
> in any case is a non-goal :-) Time to move on to something else (I
> stopped working in this area in 1991 for exactly this reason).
Again, I didn't mean to say that we can rely on end-system based CC,
and Matt and Bob have made arguments similar to yours (i.e., that we
need something in the network). My point was that what we currently
have (due to VJ) isn't completely useless (yet.)
More information about the Iccrg