[ejabberd] CPU Utilization on ejabber servers

Tiaan tiaanh at gmail.com
Fri Nov 4 10:04:03 MSK 2005


*So why don't you post an entry to http://ejabberd.jabber.ru to explain how
to do NLB? Preferably without lots of text but just with a few screenshots,
to get people started -- but that's just how I'd imagine it.*
 I will do this soon.

*If you hire a really expensive, certified Windows engineer, he'll tweak it
to make it handle more connections etc, and use less CPU.*
 I guess I'm such a person:) Windows have bee tweaked, etc, but I believe
this is a Erlang memory startup parameter, and the same will happen on any
OS with the same activity. The kernel poll not available in Windows might be
a reason moving to BSD, or like.
 Thanks for your comments.
 On 11/3/05, Andreas van Cranenburgh <andreas at unstable.nl> wrote:
>
> The following is a couple of hints & opinions. I'm trying not to
> advocate things, but only trying to point out possible room for
> improvement :)
>
>
> On Thu, Nov 03, 2005 at 08:22:56PM +0200, Tiaan wrote:
> >> 2. user activity / behaviour. are they sending each other a lot of
> >> messages? or just working and sending 1 message every hour?
> >
> > There is a lot of login / logoff activity and messages is being sent
> very
> > frequently, we also run many chat bots on the system, causing messages
> > originating from one user to be delevired to many users.
>
> frequent login / logoff: educate users to keep the jabber client in the
> systray. educate them to disable popups when the user has status "busy"
> / "away" (eg. Gajim has such settings).
>
> many bots: consider rewriting bots to use PubSub, which can let people
> subscribe to things, and then send new events to any user who is
> subscribed, with a multicast-like idea (optimised for one-to-many).
> downside of pubsub: even though ejabberd has pubsub, it might be
> incomplete (someone explain to me...), furthermore clients need to
> support pubsub (I think, again...).
>
> >> 3. if there's memory free, then you installed too much memory. Or: you
> >> can store more things in memory instead of on-disk (note that this can
> >> be dangerous in case of power-loss, data can disappear).
> >
> > Most of the Mnesia tables is configured to be stored on disk and
> > memory, the 2nd node is set to the defaults as on the first node.
> >
>
> > [snip]
>
> > Round-robin howto is still missing in both the guide and the website
> > We use Windows's NLB (Network Load Balancing) to balance the load, which
> > is working very good.
>
> So why don't you post an entry to http://ejabberd.jabber.ru to explain
> how to do NLB? Preferably without lots of text but just with a few
> screenshots, to get people started -- but that's just how I'd imagine
> it.
>
> BTW: round-robin comes down to two little DNS records. But before I'd
> write a guide I wanna test it, and see if it works properly. I need said
> server, for that.
>
> > Something else, it seems like the erlang prosses is configured not to
> use
> > more that 1GB of memory, as it's swapping while physical memory is still
> > available. Looking at the manual it seems like the "+A" and "+h"
> parameter
> > sets pool and heap values, but I don't know what the default is, nor
> what
> > to set it to.
>
> Dude, that's exactly why I'd say try UNIX / Linux.
>
> Windows is NOT optimised in its default configuration, because that is a
> way to make money for MS and the rest of the closed-world food-chain. If
> you hire a really expensive, certified Windows engineer, he'll tweak it
> to make it handle more connections etc, and use less CPU. Please consult
> the Google if you think I'm just a lying freak.
>
> The same effect (possibly better), can be attained by installing
> UNIX/Linux. Any distro would do. What's holding you back? Maybe your
> company wants to hire me? Just kidding :P
>
> The problem with Windows, I assume, is that it thinks it's pretty smart
> at deciding how to do Virtual Memory. If it's swapping, then it's
> certainly stupid, in your case! If it's wasting memory, and decides to
> swap instead, well that sounds rather braindead.
>
> But before I accuse teh venerable windows, maybe you configured ejabberd
> to do some stuff on-disk, and maybe that's the disk activity in
> question, or are you sure you're talking about lots of swapping? If you
> generate too much page faults you can get into a downward spiral of
> paging . . .
>
> Seriously, I can't help you with tweaking Windows, not even for a 250
> euros per hour I wouldn't (want to) know!
>
> [ more, anyone? :P ]
>
> --
> Andreas [ http://unstable.nl | xmpp:andreas at unstable.nl ]
> [ callto:ils.seconix.com/andreas at unstable.nl<http://ils.seconix.com/andreas@unstable.nl>]
>
>
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.4.1 (GNU/Linux)
>
> iD8DBQFDan2eAcXr2MrY86MRAl8GAJ9LMuQBw2JQe4nPw4vVnN23q3iwvACfY8nB
> kUHBVQvpaxh0/RSHYmJgE04=
> =kSF3
> -----END PGP SIGNATURE-----
>
>
> _______________________________________________
> ejabberd mailing list
> ejabberd at jabber.ru
> http://lists.jabber.ru/mailman/listinfo/ejabberd
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jabber.ru/pipermail/ejabberd/attachments/20051104/aae2cdf1/attachment.html


More information about the ejabberd mailing list