[Gllug] web server performance

Daniel P. Berrange dan at berrange.com
Thu Apr 28 13:27:37 UTC 2005


On Thu, Apr 28, 2005 at 02:17:56PM +0100, Jason Clifford wrote:
> On Thu, 28 Apr 2005, Alain Williams wrote:
> 
> > Can anyone offer tips on monitoring web server performance ?
> > 
> > I am helping to look after a small web server (apache/php/mysql) on a small box
> > (64MB bytemark - running debian). It works well most of the time, nice & fast,
> > but occasionally some users report very sluggish response. It is running at
> > a couple of hits a second, the database part is very light.
> > 
> > Looking at the apache log doesn't really help since that shows when apache sent the
> > reply to the browser - it does not show how long it took between when the request
> > arrived and when the reply was sent.
> 
> I suspect the problem is not anything you can manage directly as your 
> server is a UML instance and they are subject to io latency - particularly 
> disk io - which causes exactly the kind of thing you've indicated.
> 
> I've seen this a little here at UKFSN and it's convinced me that I need to 
> check out Xen rather than UML for the vserver offerings here.

While Xen would probably help to some degree, I'm not entirely that on its
own it would solve problems such as I/O latency, since it doesn't do all that
much in the way of resource management. For that you'd want to use something
like CKRM in your host kenrels to allow strict partitioning (or spike limiting) 
of I/O bandwidth CPU time, network bandwith, etc

  http://ckrm.sourceforge.net/


> You might be able to reduce the impact of the problem with more memory. Do 
> ensure that your memory allocation does not include swap from the host 
> system but is based only on real memory. I believe that this is the case 
> at Bytemark.

Yeah, the bytemarks hosts run with a patch to lock the memory of the UML guests
into phyiscal RAM, since kernel VMs get really upset when what they think is
RAM gets swapped out under their feet :-) I've also noticed that since Bytemark
switched all their host OS to the 2.6 kernel, latency is significantly
improved over the old 2.4 host kernels. But without any hard resource
management you still get a hit when other user's UML instances load up the
machine.

Dan.
-- 
|=-            GPG key: http://www.berrange.com/~dan/gpgkey.txt       -=|
|=-       Perl modules: http://search.cpan.org/~danberr/              -=|
|=-           Projects: http://freshmeat.net/~danielpb/               -=|
|=-   berrange at redhat.com  -  Daniel Berrange  -  dan at berrange.com    -=|
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 240 bytes
Desc: not available
URL: <http://mailman.lug.org.uk/pipermail/gllug/attachments/20050428/9faa9a24/attachment.pgp>
-------------- next part --------------
-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug


More information about the GLLUG mailing list