Re: wierd behaviour...

Kevin Littlejohn (darius@darius.wantree.com.au)
Fri, 28 Feb 1997 16:34:19 +0800


Better option might be to put the "no_atime" and "inum" patches in place
(http://www.ecsnet.com/) - these produced a definate performance
improvement for our news server.

KevinL (now why didn't I just phone you with that info? :)
Oh, btw:

root 1 0.5 0.1 832 252 ? S Feb 26 14:23 init
news 376 3.1 9.9 13700 12748 ? S Feb 26 82:51 /usr/sbin/innd -p4

Admittedly, I took the machine down the other day, but the figures still look
a little different...

> hpa@transmeta.com (H. Peter Anvin) writes:
> > By author: "Michael O'Reilly" <michael@metal.iinet.net.au>
> > >
> > > There's a simple reason for all this. The Sysvinit does a sync() in
> > > it's main loop. So every time it gets a SIGCHLD from a process w/o a
> > > parent dying, it does a sync(), which flushes all the dirty buffers
> > > (including atime mods), and voilo! Tons of disk writing.
> > >
> >
> > WHY???
> >
> > -hpa
>
> No idea. :) Belive me, next reboot that's one patch that's going in
> fast.
>
> news 8869 2.9 5.6 12568 10840 ? S Feb 21 296:18 /usr/sbin/innd
> root 1 7.3 0.0 804 48 ? S Feb 20 851:32 init [3]
>
> Init has used nearly 3 times as much CPU as innd. And that innd is
> taking a full feed! As far as I can see, it's because init is
> constantly sync()'ing producing constant runs through the disk queue.
>
> Michael.
>

---
Kevin Littlejohn                                     darius@wantree.com.au
Wantree Development                                       tel:    481 4433
Perth, Western Australia 6000                             fax:    481 0393
"Hours of frustration punctuated by moments of sheer terror" - a.s.r.