[c-nsp] MLS and accelerated switching

[Rick Ernst]

I'm still working on developing a network design for our ethernet core to
best balance the cost/value of "just moving bits".  The core is currently
a pair of 7507/RSP16/GEIP+ routers running as BGP route-reflectors between
the border and aggregation layers.

The 7507s (and GEIPs) don't have the horsepower to move much more than
about 400Mbs each with current ACLs, NetFlow, and BGP.  If the processing
were to move to an MLS or accelerated fabric, with just the high-touch
bits touching the RSP, it seems like there is still a lot of performance
available without going to an "overpowered" 7600/Sup720.

It looks like the 6500 Sup-2 supports 128K MLS entries.  Based on my
NetFlow analysis, I get the following breakdown of unique IPs per time
period:

Time    Unique IPs
-----   -----
15min   320K
 5min   150K
90sec    90K
45sec    70K
30sec    55K
15sec    35K

If I understand MLS and aging correctly, I should be able to set MLS aging
to 45 seconds and MLS flow to destination, and have at least some room for
growth.

Am I interpreting my data correctly and understanding MLS properly?  Will
MLS churn at such a short interval cause its own problems?

To alter the question slightly; is there a switching platform that could
use the RSP16s as a router-on-a-stick to handle >= 1Gbs/2Mpps?

Thanks,