Performance improvement for NAT in IPFIREWALL

[Barney Wolff]

On Wed, Jul 02, 2003 at 10:31:10AM -0700, Michael Sierchio wrote:
> 
> Currently, performance w/divert sockets and natd in ipfirewall
> on a compute-bound platform (ELAN-133MHz) shows ipfw+natd throughput
> to be 50% of that offered by ipfilter+ipnat.
> 
> Is there anything that can be done to speed up either the
> performance of divert and natd?  Alternatively, could network
> address translation be merged into ipfirewall?
> 
> As we move from 1000BASE-TX to 10000BASE-TX, this will become
> more of an issue, even on 3GHz machines.
> 
> Comments? Suggestions? Vision?

NAT is not a security feature, and should be used only where it is
actually necessary to translate addresses, and as far towards the edge
as possible.  If you believe you need to NAT at even 1Gb, I'd look
very hard at the requirements.

The performance hit on crossing the kernel-userspace boundary for natd
is inherent, apart from any code optimization that might be possible.
But moving NAT into the kernel has great impact on kernel memory usage,
which needs much more care than in user space.  NATs can be DoS'd,
and running out of kernel memory can be fatal.

-- 
Barney Wolff         http://www.databus.com/bwresume.pdf
I'm available by contract or FT, in the NYC metro area or via the 'Net.