[c-nsp] full routing table / provider-class chassis

[Gert Doering]

Hi,

On Wed, Jun 10, 2009 at 05:58:04PM -0700, Jo Rhett wrote:
> Unfortunately, Cisco's partners are useless.  They propose 6509s  
> without the DFCs, which we know will fall over.   

Whether or not you need DFCs really depends on the throughput on the
box, and the features used.  DFCs are good due to local switching (less
load on the Sup and the fabrich) and because they do local netflow - but
if the aggregate throughput is lower than what the Supervisor('s hardware
forwarding engine) can handle, a DFC will not be mandatory.

Some of our peering/uplink routers have DFCs, others have not, and with
the load we have (peak traffic ~ 4-5 Gbit/s on those boxes) the DFCs are
not yet really needed.

> And as I understand  
> it, the 6509 even with the 3CXL cards can't handle 5 full peers,  

"XL" or "non-XL" has nothing to do with the number of *peers*.

"XL" decides on the number of prefixes that you can have in your 
forwarding table (hardware FIB) - and this will be about the same 
for "1 peer with a full BGP Table" or "20 peers with the same set of
prefixes but just different BGP paths".

A higher number of different "full table "peers is going to eat up CPU 
memory and CPU power - memory is easy (Sup720-3CXL comes with 1Gbyte
RAM, which is sufficient for at least 10 "full table" BGP peers), but
CPU might reach its limit with 5 full table peers and 91 others.

Our most loaded box has 2 full table eBGP peerings + iBGP full mesh + 
~30 smaller eBGP peerings, and the CPU load is usually well below 10% -
so it might work or it might not.


> nevermind 96 total peers.   Most people suggest the 7600 platform, but  
> at least two comments on the mailing list indicate it isn't much better.

For the 7600, there is the RSP720 supervisor board, which has a faster
CPU, so it should scale better with the number


> What are people using today for this kind of environment?  Does it work?

We use 6500s with Sup720-10G (-3CXL) and Sup720-3B, and we're quite happy
with them.  The platform has its limits (shared VLAN space being the 
most significant for many folks), but compared to a "real router" (CRS-1)
the main advantage is that it's dirt cheap.

For us, questions like "does our 'router box' need to have large line
card memory to do nice QoS things in case our backbone lines fill up?"
(which is one of the big differences between LAN hardware and ES/CRS
cards) translates to "for the price difference, we can just double or
triple the raw capacity of our backbone, thus having no congestion, 
thus needing no QoS"...

(Yes, caveats apply.  With LAN hardware, you always have issues with
microbursts and buffering.  But ES/CRS - or Juniper - hardware is LOTS 
of extra money.)

gert

-- 
USENET is *not* the non-clickable part of WWW!
                                                           //www.muc.de/~gert/
Gert Doering - Munich, Germany                             gert at greenie.muc.de
fax: +49-89-35655025                        gert at net.informatik.tu-muenchen.de
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 304 bytes
Desc: not available
URL: <https://puck.nether.net/pipermail/cisco-nsp/attachments/20090611/8d77bc9f/attachment.bin>