BIND 8.2.1 creating lots of CPU usage for syslogd and named

[Volker Borchert]

In article <802.934288656 at kludge.mpn.cp.philips.com>, Jim Reid <jim at mpn.cp.philips.com> writes:

|> This behaviour is to be expect if you use syslog for query logging and
|> you're getting upwards of 200 queries a second. Each query causes
|> named to send a log entry via syslog() to syslogd, causing a context
|> switch. syslogd looks at the message, figures out what to do with it
|> and writes it to some file.

Some flavors of syslogd do an fsync() call after each and every write
so that chances of losing the last messages before the system finally
crashes are minimized. But fsync() is very expensive.

With some of them, this can be suppressed by inserting a '-' before
the filename in /etc/syslog.conf. With others, you can configure this
in /usr/src/usr.sbin/syslogd/syslogd.c ;-)

|>                             It gives up the CPU. Then named gets a
|> chance to run again (=> yet another process context switch) and 5ms
|> later it gets another query. So no wonder there's lots of system
|> activity, process switching and a busy syslogd process.

-- 

"I'm a doctor, not a mechanic." Dr Leonard McCoy <mccoy at ncc1701.starfleet.fed>
"I'm a mechanic, not a doctor." Volker Borchert  <bt at teknon.de>