[j-nsp] UDLD packets not traversing MX-80

[Saku Ytti]

On (2012-01-23 10:18 -0700), Ankur Mittal wrote:

> Has anyone out there experienced Cisco UDLD packets not traversing BGP signaled L2VPNs transparently. We are a service provider and have a Juniper MPLS core (MX-80s used as PE routers and MX-480 and T-640 as P routers). 

Unsure if relevant, but when I tested eline/rfc4448/eompls (LDP signalled) on
MX80 I noticed some peculiar behavior based on ethertype value:

- Ethertype 0x600 and higher are accepted.
- Ethertype 0x26-0x2a are accepted unlimited, 0x25 is ratelimited to
44kpps/22Mbps, 0x24 is ratelimited to 160pps/0.08Mbps
- Rest are dropped.

Seen here:
http://ytti.fi/ether_type.png

UDLD is not using ethernet2 but 802.3 IIRC, so it does not really have
ethertype, but instead it's used as frame size, so it would fall under 0x600,
as frame would be sub 1536B.

Other curious effect is, that if ethertype is recognized but you don't actually
carry the data ethertype suggests, you'll see minor packet loss:

At 11.36Mpps offered
  packet loss of 3470pps on 0x0800 ethertype (IPv4)
  packet loss of 264kpps on 0x8100 ethertype (VLAN)
  packet loss of 7kpps   on 0x86dd ethertype (IPv6)
  packet loss of 350pps  on 0x8847 ethertype (MPLS)
  packet loss of 350pps  on 0x8848 ethertype (MPLS)

This second problem is gone if you actually inject the expected payload.
Likely LU is spending lot of time looking for what it expects to find.

JNPR has so far not explained either problem. I've used 10.4R[12] for these
tests.
-- 
  ++ytti