Re: [tcpm] TCP tuning

[Joe Touch <[e-mail]>]

Mark Allman wrote:
>> Making TCP smarter is fine; doing so at the expense of robustness is
>> not. 
> 
> I'll agree with this and push back at the same time.  I agree that
> robustness is important and should be protected.  However, I have
> increasingly come to view things like initial RTOs and initial cwnds,
> and etc. as not buying us much robustness.  What buys us robustness is
> the reaction---any reaction---to indications there might be problems.
> These gross notions are more important than any small parameter change
> it seems to me.

I agree that changes to initial parameters are much less important than
the convergence/reaction mechanism that governs them afterwards - today.
That's because of the paradox that most of the connections are short,
but most of the bytes belong to long connections.

We do need to be careful that whatever we do doesn't overwhelm these
assumptions and change the game, however.

Regarding init cwnds, I'd like to see a reasonable level of analysis of
the impact. It's not enough to drive discussions with "we have to RIGHT
NOW!" to satisfy a (often underinformed) perception of need.

Joe