From: Andreas Grimm (grimm.andreasyahoo.com)
Date: Thu Nov 01 2007 - 09:34:35 CDT

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

Hello Wietse, that was the perfect hint. Load is now about 200 under stress (before 2 to 3 because of the stalled smptds), server answers immediately, even under that conditions. The only thing i've done, was to shut down the syslog. Synchronous log was turned off, of course. It's syslog-ng, so the paremeter is sync(0), but that doesn't helped here. I should try now to tune my syslog-ng (maybe the break is log_fifo_size) or install the normal syslog-package. Maybe that's an interesting addon for the stress readme. Thanks alot for your brilliant support! ----- Original Message ---- From: Wietse Venema <wietseporcupine.org> To: Postfix users <postfix-userspostfix.org> Sent: Thursday, November 1, 2007 1:34:00 PM Subject: Re: New document: STRESS_README This is the last normal activity logging from the anvil server: > Nov 1 10:13:03 mx-01 postfix/anvil[14283]: statistics: max connection rate 200/60s for (smtp:80.253.80.28) at Nov 1 10:04:11 > Nov 1 10:13:03 mx-01 postfix/anvil[14283]: statistics: max connection count 37 for (smtp:80.253.80.28) at Nov 1 10:04:17 > Nov 1 10:13:03 mx-01 postfix/anvil[14283]: statistics: max cache size 2003 at Nov 1 10:07:43 The above are logged within the same second, as it is should be. The anvil server does nothing else between logging these messages, so there never should be any delay between them. > Nov 1 10:14:52 mx-01 postfix/anvil[14283]: statistics: max connection rate 12/60s for (smtp:66.56.131.32) at Nov 1 10:14:25 > Nov 1 10:14:56 mx-01 postfix/anvil[14283]: statistics: max connection count 13 for (smtp:220.224.9.191) at Nov 1 10:13:49 > Nov 1 10:14:57 mx-01 postfix/anvil[14283]: statistics: max cache size 1974 at Nov 1 10:14:15 And these are the correspondnig strace records. They confirm that the time stamps are supplied by postfix/anvil. > send(7, "<22>Nov 1 10:14:52 postfix/anvi"..., 123, MSG_NOSIGNAL) = 123 > send(7, "<22>Nov 1 10:14:56 postfix/anvi"..., 121, MSG_NOSIGNAL) = 121 > send(7, "<22>Nov 1 10:14:57 postfix/anvi"..., 92, MSG_NOSIGNAL) = 92 Note that these records are logged several seconds apart, even if the anvil server does nothing in the mean time. Something is dramatically slowing down the anvil server or the syslogd server I suspect that it is the latter. On Linux, UNIX-domain sockets are always reliable (both stream and datagram). This has serious consequences for syslogd and its clients. IMPORTANT: on Linux, Postfix/syslogd performance sucks horribly under load, unless you put '-' before the /var/log/maillog name. Be sure to "kill -HUP" the syslogd after editing syslogd.conf. This is especially important when Postfix is experiencing a high load (either on the incoming side or while delivering mail). Wietse __________________________________________________ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com

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