Hi all,

I did some performance tests on gif interfaces on a NetBSD 1.6 system.  I 
created a number of IPv6 in IPv4 tunnels on a NetBSD host and took a 
throughput measurement with iperf by sending traffic to only one tunnel. 
Varying the number of configured tunnels(gif interfaces) on the machines 
provided very troubling results.

On a P3 800Mhz with 256 MB SDRAM, throughput was constant(92Mbps) from 1 
configured gif tunnel up to about 150.  Then, it went down following a 
logarithmic curve.  At 1000 configured tunnels, throughput was 12.8Mbps and 
at 5000, it was down to 2.04Mbps!  At the same time, the IPv4 traffic and 
the native IPv6 traffic was unaffected.  The same test on a more powerfull 
machine (P4 2.4Ghz, 512 MB DDR) gave the same curve but it started at 300 
gif interfaces instead of 150.

Activating the profiling showed that encapsulation related function took 
most of the cpu time. I attached a profiling output from a test with 10000 
gifs (just to emplify the problems) forwarding IPv6 traffic from a gif 
interface to a native IPv6 link.

Any idea on what the problem could be.  Any known limitations to the 
implementation?

Thank's for your help

Frederick Lefebvre
frederick.lefebvre@hexago.com