[j-nsp] TCAM full on EX8200?

[Pavel Lunin]

> Yes you need to look into the packet a little bit to hash well, but this
> isn't a difficult operation either (compared to holding a full table and
> doing longest prefix lookups at any rate).


As far as I understand, it's not really correct to compare difficulty of
these two operations, since they are performed by two different units inside
the chip. While label lookup instead of full IP table can dramatically
simplify the lookup unit's life, the unit, which inspects packets and
extracts bits from them, must be still quite complex even for label-only
router. Hashing ALU's life is not a peace of cake either. Say, EX series PFE
use only 6-bit seeds to construct hash on them. In case you want to push a
whole 20-bit label to the hash seed, I'm afraid, you'll need more bits in
ALU registers, more cycles or something else.

Add here the need to perform double label lookups and push/pop labels for
things like facility protection, and you'll have not that simple PFE, not
that simple software to work with it, and not that cheap product overall.

On the other hand, yes, Juniper could throw all that features away, just put
back ISIS-TE, LDP and few other MPLS control plane things back to JUNOS for
EX series and have a cheap and cheerful LSRs for BGP-free core. But,
Richard, I can't even imagine the words you'd use, if someone tried to sell
you such a product ;)

FWIW EX8200 is actually kinda bad at multipath hashing too.


I'm more or less sure, so is any cheap PFE used in ethernet switches.

I've heard (please correct me if I'm wrong), that the $1 per bucket ASICs,
used in switches, are VLIW, which is hard to reprogram. While the more
expensive ones, custom developed for routers, are rather sort of more
flexible tiny MIMD computers with asynchronously working units inside.