[j-nsp] 802.3ad trouble

[Jorn Ekkelenkamp]

Hi,

We've recently created a 802.3ad trunk between our Nortel Passport
8100-series switch and our M40.
The ae0-interface on the M40 has about 160 prefixes in the inet-family.  The
trunk consists of 3 FE-links.
All seems to be working fine at first, but after a few minutes, all hell
breaks loose, and random nexthops will
be missing in the PFE.

Our logs will be flooded with stuff like this:
Aug 11 16:38:57  dtc-core-02 scb NH: Non-existant NH (17058:Unicast, 2) in
generic change path
Aug 11 16:38:57  dtc-core-02 scb NH(nh_ucast_add): Fails to install L2
descriptor
Aug 11 16:38:57  dtc-core-02 scb last message repeated 3 times
Aug 11 16:38:57  dtc-core-02 scb NH: Aggregate nh (16953) with unknown nh in
list (17025)
Aug 11 16:38:57  dtc-core-02 scb NH: Non-existant NH (17025:Unicast, 2) in
generic change path
Aug 11 16:38:58  dtc-core-02 scb NH(nh_ucast_add): Fails to install L2
descriptor
Aug 11 16:38:58  dtc-core-02 /kernel: RT_PFE: NH IPC op 1 (ADD NEXTHOP)
failed, err 1 (Unknown)
Aug 11 16:38:58  dtc-core-02 /kernel: RT_PFE: NH details: idx 17174 type 2
ifl 135
Aug 11 16:38:58  dtc-core-02 /kernel: RT_PFE: NH IPC op 1 (ADD NEXTHOP)
failed, err 1 (Unknown)
Aug 11 16:38:58  dtc-core-02 /kernel: RT_PFE: NH details: idx 17179 type 2
ifl 175

When I disable the trunk & rename the ae0 interface to fe-0/0/0, everything
is fine, so I expected this to be a problem
with the aggregated ethernet code in de JunOS version we were running at the
time. (6.2R1.5)

In the release notes of 6.3R1.4 I found the following:
"On aggregated Ethernet or aggregated SONET interfaces, if a logical
interface has hundreds of next hops, the installation of these routes into
the Packet Forwarding Engine might take several minutes and cause scheduler
slips in the routing protocol process (rpd). [PR/48856]"

So I decided to upgrade the M40, hoping this would be a fix to our
problem.... but it wasn't.

If anybody has a clue what's going wrong here, I'd really like to know :-)

Thanks,

-- Jorn