[e2e] tcp in high rate network

[Jon Crowcroft]

I wonder if perhaps the questioner is thinking about what one might
call the
TCP equation potential well....

for a given medium, one has a pretty much non negotriale loss rate -
as you clock it faster, then the time to get to full rate for a long
loved TCP connection gets potentially longer than the life of the
typical elephant connection for the lone user on an idleish link, as
access links go faster too, this gets to be a real limit - actually 2
limits - the slow start one and the AIMD one....

we're a long way from that in most nets coz the core net is scaled
usually as the some fixed fraction of sum of the access links
and there are usually lots of conenctions, but in some research nets
where we're adding gigE at the edges and 10G in the core, there are
light load times when this is already an issue....and will go on that
way

solutions (which still fit the TCP basic CPU cost below) are not
beyond imagining....

In message <200206171245.g5HCjxmo015320 at central.cis.upenn.edu>, "Jonathan M. Sm
ith" typed:

 >>In the absence of errors and congestion TCP can run as fast as IP (minus
 >>some processing overhead) if properly tuned and presuming the end-point
 >>device has the horesepower to drive it. In practice, these latter two=20
 >>are coupled (e.g., choice of buffer size vs. host I/O capacity vs. DRAM
 >>speeds vs. slow-start vs. RTT ....). For an architectural prespective,
 >>see:
 >>
 >>
 >>%T An Analysis of TCP Processing Overhead
 >>%A David D. Clark
 >>%A Van Jacobson
 >>%A John Romkey
 >>%A Howard Salwen
 >>%J IEEE Communications Magazine
 >>%V 27
 >>%N 6
 >>%D June 1989
 >>%P 23-29
 >>
 >>For an experimental perspective, we filled most of an OC-3c between Phila=
 >>delphia
 >>and Morristown NJ with TCP/IP/ATM/SONET in 1994 - the limitations were in=
 >> the
 >>hosts, not the protocol. In later experiments in 1995, with different hos=
 >>ts,
 >>we were able to sustain 215+ Mbps, TCP/IP/ATM/SONET, albeit not over the =
 >>WAN.
 >>The limit in this case was again the host, in particular a bus in the dat=
 >>a
 >>transfer path.=20
 >>
 >>In summary, TCP seems to be able to run as fast as links as long as the f=
 >>actors
 >>pointed out above are addressed.
 >>
 >>								-JMS
 >>
 >>> Does anyone know whether TCP can run efficiently on high rate links suc=
 >>h as Tb ethernet.if the rate of access network is high too , what will ha=
 >>ppen,I am afraid that the efficiency will reduced due to the special cong=
 >>estion control of TCP.
 >>>=20
 >>>=20
 >>>=20
 >>> ______________________________________
 >>>=20
 >>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
 >>=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
 >>=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
 >>> =D0=C2=C0=CB=C3=E2=B7=D1=B5=E7=D7=D3=D3=CA=CF=E4 (http://mail.sina.com.=
 >>cn)
 >>> =D0=C2=C0=CB=B7=D6=C0=E0=D0=C5=CF=A2=A3=BA=B6=FE=CA=D6=CA=D0=B3=A1=D7=DF=
 >>=D2=BB=D7=DF=A3=AC=B8=C3=B3=F6=CA=D6=CA=B1=BE=CD=B3=F6=CA=D6=A3=A1 (http:=
 >>//classad.sina.com.cn/2shou/)
 >>
 >>

 cheers

   jon