Re: TCP

[Fred Baker <[e-mail]>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1


On Dec 15, 2007, at 2:17 AM, Jeyasekar Antony wrote:

> hi
> I heard that TCP is not suitable for high speed network because of  
> its instability, lattency.
>
> is it true? is there any research work going on in this context?

It is probably worth looking into the so-called "LAN Speed Records"  
and talking with those who have achieved them. An example of a news  
report is at http://www.itworld.com/Net/1746/030317dataspeed/, and I  
found numerous other references in the same Google search. In 2003,  
which is when that report indicated that the researchers had moved a  
terabyte of data from the US to Europe (~100 ms RTT) at an average  
rate of 2.38 GBPS, most of the backbones in use were 2.48 GBPS, so an  
average rate of 2.38 GBPS is ~96% of nominal capacity, and  
considering the overheads of TCP, IP, POS, and SONET/SDH framing,  
closer to 97 or 98% of addressable capacity.

The question is what was necessary to achieve this. They didn't use  
stock stacks - they carefully tuned a set of systems to use window  
scaling and SACK, to ensure that database accesses didn't slow things  
down, and so on and so forth. They also chose or arranged a time when  
there was little if any competing traffic.

Operationally, the guys who worry about this sort of thing the most  
are probably the astronomers, who routinely move sensor data from  
radio-telescopes across the research backbones for data reduction. In  
their cases, the sensors routinely generate in excess of 1 GBPS of  
data (64 MHz symbol rate, doubled by Nyquist sampling and again by  
channel count, with two data bits and two polarities, yields 1.024  
GBPS), and could generate any multiple of that that pleases you. They  
have been using TCP, and found that running TCP over random backbones  
and routes had a lot of problems, most notably that 1.024 MBPS is  
faster than gigabit Ethernet and TCP's loss behavior takes it  
downhill from there. Recent experiments, as I understand the matter,  
are using DCCP. They do so not because TCP is bad, but because  
offered load exceeds capacity they simply can't get all of their data  
through, and DCCP is a transport designed for that case.
-----BEGIN PGP SIGNATURE-----

iD8DBQFHZs49bjEdbHIsm0MRApDCAJ9ovo9rOEM4CuBo7tQYNPAqoQnuxQCfYLnl
99xwAeLXXZIfmGdoKYvCGqo=
=SmGW
-----END PGP SIGNATURE-----

_______________________________________________
Ietf mailing list
[e-mail]
https://www1.ietf.org/mailman/listinfo/ietf