[c-nsp] 7505, OC3s and BGP

[Matthew Crocker]

>
>> I just did a performance test with a GEIP+ and could get 400 Mbps  
>> through
>> it with 1000 byte packets. Just a data point since I tested it  
>> yesterday.
>> For the umteenth hundredth time. That card was never designed to do
>> GIG to GIG traffic. It was simply a physical layer gig *migration*
>> for aggregation router uplinks as customer migrated their cores to  
>> gig links
>>
>
> Right, but if you're running an ever growing network and have  
> outgrown FE
> for some of your trunks, you want real GigE, not a GigE card that  
> costs a
> fortune and isn't much more useful than a handful of FEs in a  
> portchannel.
> So at that point you've outgrown the 7500 platform.
>
> When I was looking at upgrading, we were able to get 6500s with  
> multiple
> GE ports and a bunch of FEs for only a few times what single GEIP 
> +'s would
> cost.

The 7500 series chassis is an OC-12 backplane (622mbps) shared bus.   
The 7513 has 2 622mbps busses and the 7507 has 1.  All packets are  
switched from VIP to VIP by the ingress VIP but the packet passes  
through MEMD on the RSP.  The RSP handles BGP and 'boils down' the  
routing information into the FIB.  The FIB is downloaded along with  
IOS to each VIP.  The VIP needs to have enough memory to handle IOS  
and FIB in order to dCEF to work.  If the VIP runs out of RAM it will  
shutdown CEF and start punting packets to the RSP. The RSP will then  
CEF switch the packets using its CPU.  I couldn't fit a full view  
(160k routes) with IOS into a VIP4-50 and  the lack of dCEF caused my  
RSP to melt.   A GigE interface in a 7500 series is GigE on the front  
and 622mbps on the back.  Using large packets you'll max it out at  
600mbps assuming nothing else is using the bus.  Using small packets  
you'll melt the VIP way before 600mbps.

The 12000 series chassis is either OC-48 (2.5gbps) or OC-192 (10gbps)  
backplane on a switched bus (2.5g, 10g per card slot). All packets  
are switched from LC to LC by the ingress LC and the packets pass  
through the Switch Fabric Card (SFC).  The GRP/PRP handles BGP and  
'boils down' the routing information into the FIB.  The 12k can  
*only* do dCEF. The GRP/PRP *will not* route packets using *any*  
other forwarding path.  If your LC doesn't have enough RAM to hold  
IOS and FIB it shuts down CEF and *stops routing*

IMHO,  512 MB RAM for RP and 256MB RAM for LC is the minimum for  
handling full BGP views.


--
Matthew S. Crocker
Vice President
Crocker Communications, Inc.
Internet Division
PO BOX 710
Greenfield, MA 01302-0710
http://www.crocker.com