On Fri, Oct 13, 2006 at 05:16:05PM +0200, Toni Mueller wrote:
> Hi Claudio,
> 
> first, I'd like to thank you for your comment.
> 
> On Fri, 13.10.2006 at 16:00:55 +0200, Claudio Jeker <[EMAIL PROTECTED]> wrote:
> > 500kpps sustained is a crazy amount of packets (especially think about
> > possible peaks). Currently you can fine tune a OpenBSD box to do over
> > 450kpps but there is not much headroom left for peaks.
> 
> Well, before specifying that packet rate, I skimmed the performance
> figures of 7206VXRs which can be made to go up to 2Mpps (using NPE-G2),
> and this gear is afair rated for a few 100MBit/s. So... when attempting
> to size such stuff, I wanted to make sure that the box holds up in case
> of DDoS and (eg.) not crash due to overload.
> 

Cisco can do 2Mpps on the G2 only in some cases (e.g. you only use the
gigabit interfaces and no acls) but honestly the NPE-G2 is currently out
of reach for any of BSDs.

> > It is better to split the load on two routers that do 250kpps each.
> 
> Erm, how do I do that on a single line?!?
> 

You can't.

> > Additionally get a fast single CPU i386 (I would use a AMD Opteron in i386
> > mode) and good network cards. This currently gives you the best bang for
> > the bucks.
> 
> Is there anything wrong with using an Opteron chip in amd64 mode?

Yes. There is a amd64 specific "bug" hidden somewhere deep down in lowcore
that caused my box to saturate at 80kpps instead of 480kpps.
I tested it about one year ago but I don't think the situation changed
dramatically.

> Wrt. network cards, I think I'm looking at bge or sk cards unless you
> want to suggest something else.
> 

I tested em(4) and bge(4) both did fine. I was not able to test sk(4) or
msk(4) (I don't own such cards).

> > Btw. 500kpps traffic as seen on the net is more than 3Gbps.
> 
> Maybe, but it depends on your traffic characteristic... If it's only
> web surfing, FTP or email, then I tend to agree. Ok, I relax to
> 200kpps, but it needs to do a little pf, carp, and a few BGP sessions
> (full table).
> 

You need to test it your self. That's why you have a testlab to gauge your
systems. Especially the impact of pf(4) depends on the ruleset, carp and
bgpd should not matter.

-- 
:wq Claudio

Reply via email to