On Thu, Jan 31, 2008 at 02:28:30PM -0500, Anne & Lynn Wheeler wrote:

> TCP requires minimum of seven message exchange for reliable transport
> .... VMTP (rfc 1045) got that down to minimum of five messages, and XTP 
> then
> got it down to three messages minimum for reliable transport (disclaimer
> we were on the XTP technical advisory board).

SMTP does not need TCP to provide reliability for the tail of the session,
the application-level "." (end-of-data) and server "250" response complete
a transaction, everything after that is optional, so for example Postfix
will send (when PIPELINING).

        DATA <CRLF>             354 Go ahead
        Message-Content         Lots of acks
        .<CRLF>QUIT<CRLF>       250 Ok

and will disconnect after reading the "250 response" without waiting
for the 221 response. The TCP 3-way shutdown (FIN, FIN-ACK, ACK) happens
in the kernel in the background, the SMTP server and client are by that
point handling different connections. So the reliable shutdown latency
is of no consequence for application throughput.

A pipelined SMTP delivery can be completed over TCP in 5 RTTs not 7.

        0. SYN                  SYN-ACK
        1. ACK                  220
        2. EHLO                 250
        3. MAIL RCPT DATA       250 250 354
        4. MSG . QUIT           250 221
        5. close socket

TCP is fine, latency is primarily the result of application protocol
details, not TCP overhead. The only TCP overhead above is 1 extra RTT
for the connection setup. Everything else is SMTP not TCP, and running
SMTP over UDP (with ideal conditions and no lost packets, ...) would
save just 1 RTT.


The Cryptography Mailing List
Unsubscribe by sending "unsubscribe cryptography" to [EMAIL PROTECTED]

Reply via email to