Other people filter at close, or better than a gig with multiple NICs.

With GigE NICs you are generally talking quality, and with quality
NICs you get lots of freebie stuff - checksum generation, decent
buffer sizes, efficient bus use etc. They know that their customers
don't want cheap, they want fast - they don't want the backups to
impact the customer experience for 2 hours, just for a measly extra $50.
Recommendations:
- Intel or other 'quality' manu.
- dedicated 66mhz bus, if possible.

CPU speed has little impact, though the little time that the frame
is actually inside OpenBSD will be halved by doubling the clock speed.
The thing is, PF is designed so that's not a large proportion of the
time - if the NIC is any good.

A search for "pf gigE Henning" is likely to bring up some interesting
stuff. e.g. http://www.benzedrine.cx/pf/msg03147.html that has figs
like:
  > No firewall:    939 Mbits/sec throughput
  > Firewall:       785 Mbits/sec throughput

But this 'problem' (17% PF slowdown) has plenty of possible solutions,
including changing both NICs to Intel, using the device polling patch,
etc.

As so many factors count, it's prob best to get a test/loan box and then
see what vmstat says while it's busy.

Dom
PS. "particularly DDoS-prone service" == online gambling ?
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Dom De Vitto                                       Tel. 07855 805 271
http://www.devitto.com                         mailto:[EMAIL PROTECTED]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
-----Original Message-----
From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of
ben
Sent: Tuesday, December 16, 2003 5:38 PM
To: [EMAIL PROTECTED]
Subject: best hardware for pf

Hi,

Anyone have experience with filtering at close to 1Gbit using pf?  What
would end up being the limiting factor in such a system?  CPU, bus bandwidth
or maybe something else?

I'm looking to take some filtering load off of our routers for a
particularly DDoS-prone service.

Thanks,
ben.


Reply via email to