[j-nsp] 8.x Compatibility (was: Upgrade without PCMCIA card?)

[Kevin Day]

On Dec 20, 2007, at 6:01 AM, Eric Van Tol wrote:

>> Downgrading to 7.6 worked for me, which is probably your best
>> bet on the M5/10/20/40 unless you're willing to upgrade the
>> RE's flash and the FEB/SSB SDRAM in a warranty-voiding kinda way.
>
> Can someone post some more info regarding this, perhaps with memory
> requirements per platform?  Will 8.5 cause problems on an M20 with an
> RE-333-768 and SSB-E?  What about an RE-600-2048 w/ SSB-E-16?  This
> seems pretty scary to me and is getting into Cisco territory now with
> having to worry about what platform can run what JUNOS version.
>

I have no idea what is or isn't actually supposed to work, but I can  
sum up my experience so far...

Upgrading to 8.5 was impossible on any of our routers because all of  
our REs had only 128MB of flash. This is easily upgradable, I just  
haven't done it yet.

We had an M10 running 7.2. 64MB FEB. RE-333-768. Full v4+v6 table plus  
RPF. After upgrading to 8.2 it failed badly, with not enough FEB  
memory. Firewall rules didn't get loaded, several thousand routes got  
dropped (trying to reach any of them would send back an ICMP  
unreachable message), etc. Turning off RPF got us back down to about  
4% free FEB SDRAM. Downgrading to 7.6 got us back to 24% free.

There seemed to be two problems there. First was that the amount of  
free memory left after booting the FEB on 8.2 was smaller than what  
7.2 gave us - so the heap was smaller before we even did anything.  
Then it appeared that the amount of FEB memory per route and per RPF  
entry was bigger than in 8.2, meaning I could do less with what memory  
we had.


We've also got an M20 (dual 64MB SSB-E, dual RE-333-768) that was  
running 7.6. Full v4+v6 table plus RPF. After upgrading to 8.4 I  
instantly hit this problem that's been mentioned here a few times -  
something gets bogged down and hundreds of thousands of routes get  
stuck in the "Pending" state when bringing up a BGP session with lots  
of routes. After rebooting, from the time a BGP session comes up until  
ALL the routes are actually installed in the PFE, takes about 15  
minutes now. "show bgp summary" will show something like:

Table          Tot Paths  Act Paths Suppressed    History Damp  
State    Pending
inet.0            739644     237987          0          0           
0     160787
inet6.0             2990       1010          0          0           
0          0

While rpd blocks on... whatever it's getting stuck on... you can't  
reach any of those 160787 routes. "show krt queue" shows tens/hundreds  
of thousands of routes waiting to go into the KRT, but I'm not sure if  
that's the cause or a symptom. Having sampling on seems to make the  
problem much worse, since rpd is even busier sending ASN data to  
sampled.

Even ignoring that problem, we had 2% free SDRAM on the SSBs once it  
was up and running. I was able to easily stick another 64MB DIMM in  
there and give some more headroom. It didn't fix the route  
installation problem, but it meant I've got 59% free space on the SSB  
now.


> Also - what type of memory can you use on an old, outdated FEB?
> Hypothetically, of course, because who would want to invalidate their
> warranty?  ;-)

There's some information here:

http://juniper.cluepon.net/index.php/Unofficial_hardware_upgrades

-- Kevin