On Fri, 21 Sep 2001, Paul Vallee wrote:

> No, unless that's the default...

No, it's not the default.

> # mount
> /dev/hda1 on / type ext2 (rw)
> none on /proc type proc (rw)
> /dev/hdb1 on /u01 type ext2 (rw)
> /dev/hda5 on /u03 type ext2 (rw)
> /dev/hdd1 on /u04 type ext2 (rw)
> none on /dev/pts type devpts (rw,gid=5,mode=620)
>
> Hmm...
>
>               noatime
>                      Do not update inode access times  on  this  file
> system
>                      (e.g,  for  faster  access on the news spool to speed
> up
>                      news servers).
>
> Seems interesting... So while running find, for example, this would save
> overhead? I'm willing to give this a try if that's what it would mean.

I've seen significative load decrease by using it, but it was on a Web
server (using apache). I think cyrus can benefit from it, but I'm not
sure whether it needs it or not to work properly.
stat (and thus find) do not change access time, AFAIK, but cpio does,
unless you use --reset-access-time. I suppose the --atime-preserve
option of tar serves the same purpose.  I believe things are worse if
you use the above options, since they'll cause double-writing the atime.
You need the -o noatime mount option to get rid of atime updates at all,
and then you don't need the two options above. Use strace(1) to see
what exaclty happens when you perform your backup. If it creates a lot
of write traffic (excluding backup output, of course), that explains
why load explodes. Cyrus already generates a good mix of read/write
traffic.

If you're using ext2, have a look at chattr(1). The A attribute seems
promising, but I'm not sure it's implemented (and besides, maybe you
need to set it on every file, and I'm not aware of any way to have it
already set on file creation). And maybe turn off -S if you set it
(following docs). It doesn't turn ext2 into a journaled FS, use ext3
(or one of the others) if you're really concerned about data integrity.
Otherwise, turn it off, and enjoy performance, at somewhat increased
risk if the system crashes. BTW, if that happenes, hope someone managed
to flush your disk controller buffers to plates in time - no FS feature
will do it.

(sorry I have no direct experience on other FS - and the cyrus server
I run isn't that loaded, I've never bothered to try noatime. And yes,
I've blindly followed instruction and set -S - so take all the above
as completely untested with cyrus)

.TM.
-- 
      ____/  ____/   /
     /      /       /                   Marco Colombo
    ___/  ___  /   /                  Technical Manager
   /          /   /                      ESI s.r.l.
 _____/ _____/  _/                     [EMAIL PROTECTED]


Reply via email to