[c-nsp] 6500/7600 TCAM Usage

Saku Ytti saku at ytti.fi
Fri Jun 3 03:10:25 EDT 2016


Hey Patrick,


> Now that we are planning to replace our supervisor engines
> (3BXL) with VSS capable ones (10G-3CXL) I'm pondering
> to repartition TCAM for 768k IPv4 and 128k IPv6 and
> to go back to full tables.
> Of course monitoring the usage closely. ;-)

Previous employer did that about 4 years ago, and would be above that
limit now too. But 3BXL runs out of DRAM sooner, so full tables were
removed from the boxes.

> I'm not asking for a time estimate when we will hit that
> limit. DFZ is at slightly over 600k v4 and about 30k v6,
> currently. And predictions are difficult, especially about
> the future.

Not really, BGP table growth is not random at all.
http://www.potaroo.net/ispcol/2016-01/bgp2015.html

> What puzzles me is: how do vendors go about that in
> the long run? I have been using my search engine of
> least distrust to no avail. Which platforms offer vastly
> bigger TCAMs, like at least twofold, better an order
> of magnitude?

Not all platforms use TCAMs. Lot of Juniper kit, like MX, QFX10k, PTX
use various types of DRAM solution, this makes FIB usually not your
bottleneck, search time to larger database becomes an issue too.
Most SP scale routers, MX, ASR9k, ALU^H^H^HNokia SR, Huawei NE etc can
hit several million in FIB.

> With RIRs handing out ever smaller prefixes I expect
> the IPv4 address space fragmentation to accelerate.

RIRs are mostly handing static prefix sizes now, not ever smaller. But
they are smaller than historically.

> Or can one get around those rather arbitrary hard limits
> completely? Is it possible to e.g. have a TCAM with timestamps
> associated to entries, so one can employ a TCAM as
> a route cache in LRU fashion and process-switch everything
> new/unknown?

Yes. But how to deal with that in hostile environment? There is
stealth startup doing something like above, but I'm skeptical. I know
some of them read this list, maybe they'll chime in.

> I did not yet take the time to browse individual datasheets
> of gear that is supposedly "bigger" than a 65k.
>
> Some pointers would be most welcome.

You're asking what platform you should buy? Usually people pay
thousands of dollars to consultants to figure that out. Or
organisations run hundreds of thousands costing RFP to figure that
out.
Without knowing your specs, I'd buy ALU SR, JNPR PTX or JNPR MX.

-- 
  ++ytti


More information about the cisco-nsp mailing list