[c-nsp] Over/Undersubscription on certain12000 Line cards

[Pete Templin]

Drew Weaver wrote:

> some of the interfaces that could be in said routers such as:
> 
> 3GE-GBIC-SC (engine 2) 
> 4OC12/POS-IR-SC-B(engine2)
> 
> Could suffer from a pretty iffy oversubscription issue:

How does a 4OC12 card get oversubscribed when the backplane interface is 
OC48?

> [snip] Except that in the case of the Gig-e card oversubscription would be
> even more damaging if you weren’t careful.

True.

> So I guess my question would be: In the case that it was absolutely
> impossible to upgrade to routers which support Engine 4/4+ would it
> make sense to jump from e2 to e3 at all? Or to save the money and do
> more when the “time comes”.

Plain and simple, I really, really wish I could be up at Engine 3 across 
the board.  What we have is working, but Engine 3 could eliminate a lot 
of the rules:

Engine 0 cards just love to drop packets, especially when hit with a 
microburst.

Engine 1 cards just plain suck: can't handle a microburst, can't 
priority queue, are incredibly featureless.

Engine 2 cards seem to be much better, but still don't have enough route 
memory to be "safe": we have a 12008 with OC12 links to the core (card 
redundant) and the 3GE with one port live on a transit feed to AboveNet. 
  Since AboveNet feeds us (and presumably most customers) off two 
Junipers through a Catalyst6500, we get two BGP feeds.  Enabling BGP 
multipath doubles the size of the FIB, causing all of the Engine 2 cards 
to run out of RAM.  It's OK to have both feeds, but just don't (ever) 
run BGP multipath.  And don't get me started about the inability to have 
both an input ACL (those are supposed to be easier, right?) and uRPF 
simultaneously on the card.

My original thought was to start with the E2 3GE (with a 1-port internal 
rule/limit), and upgrade to E3 4GE, moving the E2 cards into the core 
(where the ACL and uRPF limits aren't an issue), but the multipath 
limitation could/would still be an issue.  :(

pt