[e2e] TCP un-friendly congestion control
Injong Rhee
rhee at eos.ncsu.edu
Fri Jun 6 10:51:31 PDT 2003
Hi Craig,
Well.. I wasn't commenting on a particular version of TCP; I am
commenting on its inherent limitation under fast, long distance
networks.
My definition for fast, long distance networks might have been perhaps
too liberal. In case of 1Gbps and 50ms, its B x D is 4167 packets
assuming 1500 byte packets. It will take around 2083 (4167 after worst
case timeout) round trip times which is around 104 (208) seconds. If the
delay increases to 100ms and 200ms, this becomes 6 min and 28 min. As
you scale the bandwidth to 10G, this becomes ...well..I guess that
should be enough.
These numbers assume the best performing TCP (SACK, PAW and window
scaling and what else). If you say that these networks do not exist,
well... that is a different story. ESnet and Internet2 are the common
examples. If you say who uses these high bandwidth flows, well welcome
to high-performance computing.
Injong Rhee
Computer Science Dept
North Carolina State Univ.
rhee at csc.ncsu.edu
http://www.csc.ncsu.edu/faculty/rhee
-----Original Message-----
From: Craig Partridge [mailto:craig at aland.bbn.com]
Sent: Friday, June 06, 2003 1:09 PM
To: Injong Rhee
Cc: end2end-interest at postel.org
Subject: Re: [e2e] TCP un-friendly congestion control
In message <000001c32c43$6231fce0$9d350e98 at HD>, "Injong Rhee" writes:
>It is my understanding that in the environments where these protocols
>(HSTCP, Scalable TCP, and FAST) are intended to operate -- fast
(>1Gbps)
>and long distance networks (>50ms) -- TCP friendliness is not a concern
>since TCP cannot fully utilize the available network bandwidth.
TCP can achieve performance in excess 1 Gbps on terrestrial links. You
generally need to get to satellite links to have trouble and there the
maximum rate is around 1.5 Gbps over geosynchronous.
Just to be careful, as to specify TCP performance you need to say
several
things: my statement is for a TCP with PAWS, SACK, and window scaling,
over terrestrial networks (delays < ~100ms) with appropriately low loss
rates.
Craig
More information about the end2end-interest
mailing list