>I tried the suggested modification (commented out the USE_ENCAPCHECK 
>defines) and it makes a huge improvement.  I can now get about 1250 gif 
>configured gif interfaces with no performance cost.  But then, the 
>throughput goes down in the same way that it did without using radix trees. 

	what about throughput in total? (i.e. total of all the tunnels)

>I would have expected the lookup time using radis tree to be more linear. 
>But when I plot my results, it seems the curve as not changed at all... it 
>has just been move to the right.  What really puzzle be are the profiling 
>results.  I attached two profiling output. One is from a setup with 10000 
>configured v6 in v4 tunnels with NetBSD 1.6 while sending tcp traffic 
>through 1 tunnel.  The other is the same setup but with the "#define 
>USE_ENCAPCHECK" removed.  Comparing the two, we see that "encap4_lookup" is 
>a lot faster but it is called a lot more often and now take over 55% of the 
>CPU time.  Does anybody have an idea of what is going on here?

	do the two profiled results use the same amount of packet?  if the
	amount of packet forwarded was different, it could be misleading.

itojun