Re: [tcpm] TCP tuning

[Joe Touch <touch at ISI.EDU>]

Murali Bashyam wrote:
> Any idle period higher than the RTT between HTTP requests tends to
> reset the cwindow down to the restart window.

This was first documented in an extension to Jacobson's 1988 Sigcomm
paper, in an appendix. The idea is that if you are idle, you shut the
window down. However, a footnote in that doc (unpublished, FWIW)
suggests that because TCP is used mostly symmetrically, one can use the
"time since last receive" instead of "time since last send". As a
result, the request tends to reset the timers, and the server just
blasts at line rate.

This issue is documented in detail in:

John Heidemann, Katia Obraczka, and Joe Touch, "Modeling the Performance
of HTTP Over Several Transport Protocols", IEEE/ACM Transactions on
Networking, V5, N5, Oct. 1997, pp.616-630.

A correction to TCP's behavior was suggested on this list, and we worked
on it with Sally Floyd for a while ("slow start restart" was the phrase
we used). A number of variations were considered, including 'burst or
lose', 'use it or lose it', etc. We put that on the shelf a while ago,
though we've been interested in 'restarting' it too ;-)

Joe

Attachment: signature.asc
Description: OpenPGP digital signature