[e2e] TCP 'fast start'

David P. Reed dpreed at reed.com
Thu Sep 29 07:07:26 PDT 2005


Khaled Elsayed wrote:

> I think all these effort to enhance TCP are great. However, it seems 
> that TCP has become too old and inefficient for many cases (wireless, 
> short connections, ultra-high speed, etc).

Read the original TCP papers and so forth.   TCP was never "efficient" 
nor was it intended to be.   It was intended to be *interoperable*.   It 
was an *overlay* network.  It still achieves those goals far better than 
an "efficient" protocol (especially since efficiency seems to come at a 
cost - specialization, brittle response outside a narrow set of 
operating points, etc.).   That doesn't mean that TCP achieves its goals 
fully - but if you focus on goals it never had (like "efficiency") as 
your critique, you will end up with SNA (which was FAR more 
bit-efficient than TCP ever was, but failed to be useful in a broad 
context) or SS7 and the Bell System 3 kHz highly-optimized billing 
machine - also far more "efficient" than TCP but only in a narrow domain.

>
> Isn't it time that the research community should think about keeping 
> TCP interface (sockets) but come up with a more efficient universal 
> protocol or am I just too optimistic/unrealistic?

The research community should be exploring a wide range of things.   
Resilience and adaptability is far more important than "efficiency" .   
Far too many in the research community are focused on a narrow set of 
metrics, invented by academics, for academics, merely because they are 
quantitative.   They are the "drag racing" community, who focus on 
top-fuel funny cars (perhaps that's too US centric?).   We need more 
researchers focused on how the cars fit into the human ecology of 
communications, in particular some ought to be thinking about inventing 
better metrics for things like adaptability and resilience, which are 
far more relevant systems properties.   Can anyone tell me a defensible 
measure of adaptability that has been used to rank network performance 
in the real world?




More information about the end2end-interest mailing list