[e2e] TCP un-friendly congestion control

David P. Reed dpreed at reed.com
Sun Jun 8 08:15:52 PDT 2003


At 07:50 PM 6/7/2003 -0700, Steven Berson wrote:
> > Instead of running such a network, increase the capacity of the bottleneck
> > links to 10x load!  It's cheaper and the customers are happier.
>
>I don't think you have made a convincing economic argument here.

Perhaps not.  "The customers are happier" implies a whole analysis that 
focuses on what customers actually value.   The assertion that 10 customers 
getting 1/10 of the bottleneck bandwidth is the economic optimum is 
indefensible, because there is no sound economic argument that supports 
either "fairness" or the maintanence of bottlenecks despite the customers 
having money they would gladly invest in eliminating the bottleneck.

"Fairness" is an arbitrary engineering judgement that disregards the 
relative importance of flows, and the idea of keeping the bottleneck full 
disregards the relative importance of reduction of latency for certain 
types of flows.

Finally, the "achievable performance region" (which I would define as the 
range of operating points of the network routing architecture and its 
router queues as an ensemble) almost certainly has much higher value to 
customers when the bottleneck queues are not full (I would argue that the 
maximum end-user value occurs when all queues are close to empty!).  This 
is because the value to end users is not equal to bits delivered times a 
constant that is independent of the users.   I, for example, am happy to 
pay more for a network that displaces other users by *empty* capacity in 
order that at any point in time I can launch packets into the network that 
will not be delayed by traffic that other users are putting out.   Even if 
they back off quickly, they will still impact my more important traffic.

The word "economics" is often taken in vain on this list by people focused 
on micro-optimization of some component of the system.   No economist would 
confuse micro-optimization with global value maximization, but engineers 
often do. 




More information about the end2end-interest mailing list