From: Mark Martinec (Mark.Martinec+postfixijs.si)
Date: Thu Jul 26 2007 - 14:15:48 CDT

• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Wietse Venema:
> Your earlier trace shows that the receiver window is wide open
> (35840 - 4096 is > 0), yet the sender waits for an ACK anyway:
> And this is because FreeBSD lo0 has an MTU of 16384. Sending 4096
> bytes triggers a Nagle delay.
>
> TCP_NODELAY is a workaround (especially with ethernet it would
> generate more packets than necessary). A better solution uses
> TCP_CORK (Linux) or TCP_NOPUSH (*BSD) where available. All systems
> can benefit from a bigger VSTREAM buffer size, but that is not an
> option for the stable release.

Just for fun, I tried with TCP_NOPUSH:
  http://www.ijs.si/~mark/tmp/dkim-pf-nopush.tcpdump.bz2

It bumps elapsed time to 7.5 minutes(!), each wait is 5 seconds long.

> Thanks for spotting this.

thanks to a timing breakdown in amavisd-new log, which showed that
a content filter waited long seconds for a SMTP status response from
MTA, after sending the final dot on mail re-submission (5.5MB mail):

(34690-05) TIMING [total 12900 ms] -
  SMTP greeting: 3 (0%)0, SMTP EHLO: 1 (0%)0, SMTP pre-MAIL: 2 (0%)0,
  SMTP pre-DATA-flush: 4 (0%)0, SMTP DATA: 665 (5%)5,
  digest_body: 54 (0%)6, sql-enter: 24 (0%)6, mime_decode: 1158 (9%)15,
  get-file-type1: 504 (4%)19, decompose_part: 881 (7%)26,
  get-file-type1: 582 (5%)30, update_cache: 13 (0%)30,
  fwd-connect: 12 (0%)30, fwd-mail-pip: 3 (0%)30, fwd-rcpt-pip: 0 (0%)30,
  fwd-data-chkpnt: 0 (0%)30, write-header: 1 (0%)30,
  fwd-data-contents: 285 (2%)33, fwd-end-chkpnt: 8644 (67%)100, ...
                                                 ^^^^

Mark

• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]