Re: 2.2.19pre6aa1 degraded performance for me...

[Sasi Peter]

On Tue, 9 Jan 2001, Sasi Peter wrote:
> On Tue, 9 Jan 2001, Andrea Arcangeli wrote:
> > On Mon, Jan 08, 2001 at 10:46:29PM +0100, Sasi Peter wrote:
> > > What I had w/2.2.18pre19 (+raid+ide):
> > > ~80MB more in cache and ~80MB swapped out (eg. currently unused notes
> > > server and squid) There is enough of swap over 3 disks (like the
> > > raid), so I did not bother disabling squid and notes, since - I thought -
> > > they would only take up some swap unused.
> > There are many variables. However I guess the slowdown is because your idle
> > apps didn't got swapped out in favour of cache as you noticed. An aggressive
> > aging algorithm would probably fix that but it then would hurt other cases
> > (after you don't need a frequenty accessed part of filesystem cache anymore it
> > would take ages before it gets collected potentially causing an unnecessary
> > swapout storms because the kernel doesn't know you don't need such cache
> > anymore).  Furthmore if notes and squid are rarely running but they provides
> > critical services if they would go totally into swap in favour of fs cache you
> > would get very bad latencies the first time somebody connects to the server. So
> > the fix I suggest you is to buy more ram or to shutdown squid and notes. Than
> Oh well I thought 384MB should be enought for everyone aiming at this
> performance (almost TM ;). At least it would up till now :(
> > you may as well see a performance improvement compared to 2.2.18pre19
> > (+raid+ide).  Otherwise you can push the machine low on memory a bit until they
> > both goes totally into swap (check with `ps v`). Hope this helps.
> I'll try this, thanks. (so no echo '1 23 456' >/proc/sys/vm/...?)

I thought it over again. I still have to say it is a nonsense for a kernel
not to  have _anything_ (zero pages) currently unused swapped out under
such an I/O load!

-- 
SaPE - Peter, Sasi - mailto:[e-mail] - http://sape.iq.rulez.org/