[jadmin] Bouillon p2p wiki and Jabber server load

[Victor Grishchenko]

On 01.07.2006, at 0:09, JD Conley wrote:
> That is quite a bit of traffic. Why so much?

Bouillon assembles a page from pieces collected in some social  
vicinity of a user (i.e. friends, friends-of-friends). Each piece is  
processed separately, so 100pieces*10friends~1000 or  
10pieces*100friends~1000. Because of almost-real-time collaboration  
features, requests have to be refreshed every minute, so 1000 stanzas/ 
minute.

> The fastest servers out there (that I've personally seen) only  
> route somewhere in the low thousands of messages per second per  
> processor. At 17 IQ's per second per user (1000 per minute), you'd  
> max out CPU's with only a few hundred users with your product on an  
> average server. I'm also concerned about the bandwidth requirements  
> of the system.
> At this level of traffic and bandwidth I could see your application  
> getting blocked very quickly by most public server admins. Also,  
> obviously, if you're compressing stanzas in both directions CPU  
> utilization will be much greater.

May use of peer-to-peer bytestreams soften the problem? This way,  
only NAT:NAT pairs will use server facilities, right?

By the way, a question: which variant of routing traffic of NAT:NAT  
pairs is less abusive: bytestreams/SOCKS5 or in-band stanza routing?  
I see two 'pros' for the former variant: compression is done by  
clients and servers don't do routing. The 'contra' is that it needs  
many simultaneous persistent connections to the proxy (several conns  
per every natted user).



								Victor