[c-nsp] ASR100x route tables sanity check

[Łukasz Bromirski]

On 2012-02-04 17:36, Gert Doering wrote:

> On Sat, Feb 04, 2012 at 03:16:18PM +0100, ?ukasz Bromirski wrote:
>> ESP5 on the other hand can store approx. 512k IPv4 routes *OR* 128k
>> IPv6 routes. For a mix of these, rule of thumb is pretty obvious - you
>> take 4 IPv4 routes for each 1 IPv6 route installed. Please note however
>> hardware specific checks like the one found on for example 6500/7600
>> units with carving of TCAM space are not present here - the mix will
>> be dynamic.
> What happens if ESP runs out of FIB slots?

That's still a bit complicated issue. Generally, FIB uses the DRAM
space in the ESP for storing FIB entries. You can check the state
using 'show platform hardware qfp active infrastructure exmem
statistics' and look for DRAM numbers. If the DRAM runs out of space,
IRAM will be used, but that's "Instruction RAM" and from this point
going forward, different bad things can happen, up to and including
ESP crash.

As a side note, ASR1k has software forwarding path indeed, but it's
"hardware" rate-limited by the QFP. You can check the current
situation with punted traffic using following command:
"sh platform hardware qfp active infrastructure punt statistics type per-cause"


> On a 6500, the result is not pretty (falls over to software forwarding
> for some prefixes, and will never recover if the number of routes in the
> FIB falls under the threshold again).

That's more complex.

Historically, before the 6500/7600 split, the exception
for FIB overflow was to forward *all* traffic using the
RIB, so purely control-plane. That was usually killing the
box. You could however use hidden
'mls cef exception action freeze|restart|clear' to somehow
tune the behavior, but it was not officially supported.

On the 6500 after the split in the 12.2SX train, and
after the CSCsq77464, the behavior changed somewhat - if
indeed exception happens, system will autoconfigure
mls rate-limiter to 'receive' path up to 10kpps, and use that to
forward traffic using the RIB table. This is more bearable in real
world, but still far from ideal.

On the 7600 however, in the 12.2SR traing and specifically from
12.2(33)SRB3 on, the system will program drop adjacency for the
prefixes that can't fit into TCAM FIB.

HTH,
-- 
"There's no sense in being precise when |               Łukasz Bromirski
  you don't know what you're talking     |      jid:lbromirski at jabber.org
  about."               John von Neumann |    http://lukasz.bromirski.net