Richard Carlson wrote:
You're right, the TCP peak was 1.48Gbps from the show floor in Dallas to
storage cluster at Berkeley Laboratory. Single applications using
stream. Bottleneck link was 1.5Gbps provisioned circuit on Qwest link
convention center to DARPA's HSCC pop node.
This beat last years (SC'99) 1.2 Gbps rate Microsoft achieved running
on a single PC with 2 Gbit Ethernet cards (Redmond to Portland).
Peak rate alone is a meaningless measure of performance. The TCP session
have reached peak for a small fraction of the session duration and the rest
the time, the link may have been severely under-utilized.
A meaningful performance measure here is average or percentile link
over the tcp session duration.
I believe the inital question was answered, though. Peak rate alone tells
me that it is _possible_ in the "near" future to exhaust the sequence number
limitation of TCP.
Even still, is it wise to allow for such conditions? Or would it be wiser
to influence the community use larger frames over these large, fat links?
Send a cool gift with your E-Card