J2350 crash after starting iBGP session

[ehe@nbs-system.com]

Hi,

We have 2 J2350 routers, each one being connected to a different eBGP neighbour, two upstream providers. Each eBGP session give us about 327000 BGP routes.

To optimize outgoing traffic, we want to establish an iBGP session between the two J2350.

When we tried it, one of the two routers crashed, but not immediately. The session was established after a "commit confirmed", a "show bgp neighbor" showed the iBGP session to be established on both side and "show chassis routing-engine" showed a high memory usage (about 90%). Everything seemed to be running fine, so we ran the "commit check" command to confirm the commit. At this point, one of the two routers was stuck here :

imil@jun1# commit check

I was able to type [enter], the TCP session was still active on port 22, but even the serial console was not responding, we had to hard-reboot the router.

Right now, we don't really know what to think: is it a bug ? is a J2350 with 1GB memory unable to handle 1 eBGP session + 1 iBGP session ?

Thanks in advance for your replies.

Regards

[tontsa]

In J2320 when I last tried your scenario I had to install 2GB of memory. Also the sync was very slow.. took around 20 minutes for the session to sync all routes. I don't really recommend running full tables with these old J-series as they are basically P4/P4 celerons.

[rfrederick]

I would agree, a J2320 is not designed to take a full set of routes.

Ron

[iMil]

Well, this is pretty annoying since the persons who told us this would cause no problem were two Juniper engineers before we bought those J2350... and yes, we told them we were to bring up eBGP + iBGP.

Anyway, we finally brought the iBGP session up between the two, the only different thing we did was to wait a little bit more before the "commit check". Our guess is that the first time, the routers were under heavy load while establishing the iBGP session, and as we noticed a "commit check" is really CPU intensive, it probably overwhelmed the poor J2350. Good to know we must be extra-careful with those.

Thanks for your replies.

[Dominik]

Hi,

I have used also J2320 routers with 2x full feed on them. Works good. But I downgraded to the last JUNOS 9.3 that runs only in packet mode. That saves a couple of RAM.

Regards,

Dominik

[iMil]

Hi Dominik,

This is very interesting, we're running 9.6R4.4, would you suggest us to downgrade to 9.3 ?

ATM, with both eBGP and iBGP sessions being up, we have the following values :

    Total memory              1024 MB Max   922 MB used ( 90 percent)

      Control plane memory     594 MB Max   564 MB used ( 95 percent)

      Data plane memory        430 MB Max   357 MB used ( 83 percent)

    CPU utilization:

      User                       8 percent

      Real-time threads         14 percent

      Kernel                     6 percent

      Idle                      72 percent

[Dominik]

If you don't need any flow based features, I would definitly give that a try. Be carefull to install the NON enhanced security version of JUNOS to get the pure packet based version. The last release was 9.3 R3.

Regards,

Dominik

[iMil]

Looking at 1st top(1) entry, your solution makes definitely sense :

 1042 root        1  98    0   481M   481M select 396:24  1.03% flowd_hm

Thanks a lot for the hint !

[iMil]

And so we did it... and it worked !

No more flowd_hm process, much more responsive router, eBGP / iBGP sessions up.

Model: j2350

JUNOS Software Release [9.3R3.8]

[Dominik]

Hi,

great to hear, you are welcome!

If you like, Kodus are welcome

Regards,

Dominik