[e2e] What should e2e protocols know about lower layers?

[Marcel Waldvogel]

Christian,

I agree with you in the general case (long, obese pipes). It would be highly 
desirable if one of the main properties of slow-start (roughly determine the 
useable bandwidth within a few round-trip times) could be applied to 
congestion-avoidance phase as well.

But in the context of the original query ("should congestion control be turned 
off for 'local' connections?"), even with a linear increase, a typical local 
link (short, obese pipe) will be able to grab the link again within well below 
a second even with linear increase.

-Marcel

Christian Huitema wrote:

>>I agree that TCP in the congestion avoidance phase will take forever
>>to absorb the remaining bandwidth. 
>>
> 
> Now, that is a real problem. It means that any application will be
> better off with a set of short lived connections than with a single long
> lived one. Do we expect application developers to not notice that?
> 
> The mathematics are clear. If we want a solution that scales to
> arbitrary speed, that solution should not include non scalar constants
> such as "one packet per RTT", which the congestion avoidance does. The
> equivalent slow start constant is "at most double in one RTT", which is
> still non scalar but at least does not include an implicit bandwidth
> assumption.
> 
> -- Christian Huitema 
> 
> 
>