[asterisk-dev] rtp scalability improvement...

[Andrew Kohlsmith]

On Sunday 19 March 2006 16:01, Roy Sigurd Karlsbakk wrote:
> After some discussion on #kernelnewbies on irc.oftc.net, one whos
> nick is 'ahu' told me this was 'because of heavy switching', too many
> system calls etc. he mentioned using mmapped libpcap as a dirty hack
> that would allow packets to move into userspace without any system
> call at all.

I ran this by my kernel-hacker friend and he largely agreed, but suggested 
performing this dirty hack and profiling it just to see what happens.

My own thoughts are that mmaping pcap in this manner largely bypasses many of 
the things that the kernel can do to packets (which is the point) but you 
need to be careful because you can lose a lot of the functionality you may 
have been expecting if you are relying on any advanced routing or processing.  
However, I imagine that the people interested in doing this have any packet 
processing done on separate routers, so the point's moot.

To me, 400 calls on a single Xeon seems like a pretty good balance between 
scalability and performance...  this is all commodity hardware, remember.  I 
wonder, Roy, what the bandwidth and packet levels were like.  (i.e. what did 
the network look like?)

-A.