stpeter at jabber.org
Tue May 11 10:24:02 CDT 2004
On Tue, May 11, 2004 at 03:48:29PM +0200, Philipp Hancke wrote:
> > XMPP already is highly popular -- I estimate 10+ million IM users. And
> > so far the network usage is by no means extreme: on average, probably
> > about 15 bits/second for every connected user, so that a server with
> > 1,000 concurrent users would consume about 15 kbs of bandwidth, 10k
> > users would consume 150 kbs, etc.
> Just a short question... what traffic are you estimating?
> a) the traffic of an isolated server
> b) the traffic of an server that is connected to many others
The average bandwidth used by one user is, in my experience, 15 bits per
second over the life of the user's session. This is on servers that are
connected to the rest of the Jabber network, not isolated servers locked
away behind firewalls or whatever.
> Lets assume a user that has 500 roster items, 400 of them non-local,
Very very few people have 500 roster items -- I know the numbers for the
jabber.org server, and the average user probably has 10 roster items,
"power users" have perhaps 50, and there are a very few outliers on the
bell curve who have 200+ roster items -- but very very few.
> goes online. This triggers a presence packet of 200 bytes to each
> person on his roster.
Why do you think a minimal presence stanza is 200 bytes? I estimate that
it is closer to 70 bytes. But for the sake of the math I'll grant you
> 100 of them are delivered locally. Makes ~20kb.
I'd say 10k.
> 400 are delivered to remote servers. Makes ~80kb going around the
I'd say 40k.
Compared to VoIP, file transfer, video streaming, file sharing, and
email messages with large attachments, this seems rather minor, no?
I have 800+ roster items -- the most on the jabber.org server. So when I
log in, I perhaps generate 80k of bandwidth. But after this spike in
usage, I'm online for 10 hours at a time and even a popular IM target
like me is usually chatting with not more than one or two people at a
time. So my feeling is that your worries are overblown.
> Same scenario using JEP-033 and ip multicast:
> Client sends one stanza of 200 bytes to the server, which adds 500 * say
> 30 bytes per address node.. makes ~15k roughly. The server forwards
> this to an ip multicast address, thus letting 15k pass the internet
> just once. Actually less, because he can remove 100 local users..
> Or can he even leave out all addresses when using ip multicast?
> > I'm still not convinced there is a problem with the XMPP specs as they
> > are today. However, in certain situations, extended addressing (as in
> > JEP-0033) might be helpful.
> Even if the implementation of a distribution strategy is a too complex
> problem for the current specs, I find that XMPP should be ready to
> handle such an extension when introduced later...
XMPP *is* ready -- we have JEP-0033 defined and ready to go, and I think
that several server implementers have started to code to it (or they
will now that it is Draft within the JSF's community standards process).
Jabber Software Foundation
More information about the xmppwg