[j-nsp] Full routes on MX5

[Michael Loftis]

On Tue, May 10, 2016 at 1:12 PM, Adam Vitkovsky <Adam.Vitkovsky at gamma.co.uk>
wrote:

> > Michael Loftis
> > Sent: Wednesday, April 27, 2016 6:48 PM
> > To: Matthew Crocker
> > Cc: juniper-nsp at puck.nether.net
> > Subject: Re: [j-nsp] Full routes on MX5
> >
> > You'll definitely be a lot happier with the bigger RE's...usually my
> > convergence time at $dayJob is generally 3 minutes or less, with the less
> > often depending on how fast we get routes form the other guy when a
> > transit flaps.  Cold starts are a little ugly-ish with the number of
> full tables we
> > take in but still ~5 minutes usually once booted...that RARELY happens
> > though, esp now w/ 14.2's ISSU on MPCs...  This is on an MX960 w/ MPC's
> --
> > DPCs actually slowed the RIB->FIB process down I don't remember exact
> > timings sorry -- 14.2 train as well which makes huge differences if
> you're
> > using flow, and that definitely slows things down.
> >
> So do I understand it right that your experience is that 14.2 further
> slows down the RIB to FIB convergence as I'd assume the opposite?
>

No, sorry, you've misunderstood.  It's no better, no worse with same
hardware.  DPCs are slower to get the RIB->FIB convergence, with MPCs we
generally have not noticed any RIB->FIB convergence WITHOUT flow sampling
on all supported releases.  WITH sampling prior to more recent revs we DID
see RIB->FIB issues with MPCs as well.

For our use cases in production (not using sampling in production on DPCs!)
it was DPCs causing occasionally slower RIB->FIB times for us.  Never
really had any serious issues like many have had but I put that down to
having not used sampling/flow until well after Juniper had taken
significant steps to address the sync issues as being the biggest reason.
 14.2 didn't seem to change the timings except in problematic cases in the
lab (flow sampling) but we didn't use flow sampling in production prior to
14.2 at all so I can't speak from any real experience there.


>
> Also I'd like to ask if you've considered using hierarchical FIB as a
> workaround for the slow RIB to FIB convergence?
>

In our case except in testing with flow prior to 14.2 (unsure which
revision), and with a cold start we've not really observed issues with
RIB->FIB convergence in our use cases after upgrading to MPC3E's and
later.  With the prior generation setup using DPCs there were occasionally
some significant RIB->FIB sync problems but in all cases I recall it
settled within a couple minutes.


In general when I'm speaking of convergence I mean the whole thing end to
end.  Which means you're also at the mercy of your peers because if they're
not sending routes as fast as you can process them, you'll get slowed
down.  If your hardware at any point has a bottleneck you'll get slowed
down.  I'm generally NOT discussing the RIB->FIB step alone.

-- 

"Genius might be described as a supreme capacity for getting its possessors
into trouble of all kinds."
-- Samuel Butler