[c-nsp] Cisco vs. Juniper

[Richard A Steenbergen]

On Mon, Nov 02, 2009 at 05:14:30PM +0000, Mackinnon, Ian wrote:
> Not wanting to disagree with the mighty Steinar :-)
> If you have any significant amount of traffic you need to be sampling at
> over 1/1000 or you will kill the link to main cpu. Juniper and our 3rd
> party support company explicitly said "don't do it"
> 
> We had a couple of incidents where our traffic went to a full 1G and our
> 1/100 sampling totally killed the box.

It is only a 100Mbps link between the routing engine and CFEB, but I
don't think you'd be filling the port even with 1/100 sampling. You'd
certainly overload the software sampling capacity, and I suppose you
might bump a hard coded rate limit they never expected anyone to bump
(which sounds like the case, if it broke regular forwarding). Don't do
1/100 sampling and you'll be fine. :)

> Up until then, I thought if a M7i did anything, it did it at full line
> rate, always.

Actually it doesn't do line rate forwarding either. The "FPC1" component
(the 4 main PIC slots) does a peak of 3.2Gbps full duplex, before taking
into account jcell overhead (this is a limitation of access to the
packet buffer memory). Under artificial conditions (65 byte packets,
which consume 2 64-byte jcells) you can force performance down to just
under 2.5Gbps. Remember the FPC1 was originally designed for OC48s back
in 1998 when Cisco had nothing that could compete with it. It's a
testiment to the quality of the design that you can still use it for a
couple GE's under non-extreme traffic conditions today (I don't see
anyone still trying to use their 7500s to do the same :P), but obviously
it's not going to compete with modern hardware.

At any rate, this is the wrong list so I'll stop responding with Juniper 
information unless you wanna move it over to j-nsp. :)

-- 
Richard A Steenbergen <ras at e-gerbil.net>       http://www.e-gerbil.net/ras
GPG Key ID: 0xF8B12CBC (7535 7F59 8204 ED1F CC1C 53AF 4C41 5ECA F8B1 2CBC)