[j-nsp] MX80 Junos ver recommendations

[Richard A Steenbergen]

On Fri, Oct 15, 2010 at 12:38:48PM -0500, Richard A Steenbergen wrote:
> I'd suggest 10.2R3, some time in November. There are just a massive 
> number of things that are very very broken in 10.2R2 (at least in our 
> testing of the Trio cards on MX960), and no fixes planned for any of 
> them until 10.2R3.

FYI I'm told 10.2R3 is actually scheduled for Oct 18th, though with the 
way these dates tend to slip that might actually mean November anyways. 
:)

In addition to the issues I talked about previously on 10.2R2 for Trio 
(rsvp allow-fragmentation causes blackholing of packets entering LSPs, 
certain filter match criteria crashes MPCs, and mib2d crashes repeatedly 
because of a memory leak and requires a manual restart mib2d to make 
SNMP work again), we've also run into the following issues in 10.2R2 
which should be fixed in 10.2R3:

* MPLS octet counters don't work when you're doing implicit null and 
passing the packet cross the wire without a label. This is a big problem 
for ingress nodes on single hop LSPs, since auto-bw will record no 
usage, and your rsvp-te will break. A workaround it to configure 
explicit-null on the egress nodes that are one hop away from the Trio 
cards.

* SNMP sometimes returns incorrect counters for logical interfaces. We 
have one router right now where every logical interface appears to be 
counting only local control-plane destined traffic, even untagged .0 
interfaces, but the physical interface returns correct usage. We've also 
seen instances where a vlan-tagging trunk with 3 units started returning 
counters showing the physical interface usage on all 3 subinterfaces 
simultaniously. Sometimes these issues magically fix themselves, other 
times they persist, and restarting mib2d/snmpd doesn't seem to help it.

* Powering off DPCs in a mixed DPC+MPC chassis can cause the MPCs to 
crash, and stay down until the box is restarted (or in our case, we did 
a manual RE switchover to kick it).

As I said these are all MPC+MX960 issues, but the PFE on MX80 is almost 
exactly the same, and most of these are sw bugs related to the PFE, so I 
would expect a lot of the issues to translate over. The SNMP issue is 
the most obnoxious for us, since it screws with the ability to bill 
customers. If you do run Trio cards prior to 10.2R3, definitely keep an 
eye out for this issue appearing and disappearing, seemingly at random. 

-- 
Richard A Steenbergen <ras at e-gerbil.net>       http://www.e-gerbil.net/ras
GPG Key ID: 0xF8B12CBC (7535 7F59 8204 ED1F CC1C 53AF 4C41 5ECA F8B1 2CBC)