Re: [Idr] draft-litkowski-idr-bgp-timestamp

[Jeffrey Haas <[e-mail]>]

Sue,

On Thu, Jul 24, 2014 at 02:50:06PM -0400, Susan Hares wrote:
> Based on work with vendor boxes for the BMWG BGP work - I would recommend
> you timestamp the BGP trace NLRI as you leave the BGP Peer and as you
> receive it. Otherwise, you run into the data flow patterns the trace NLRI is
> included in causing variation in your flows.

We chatted further about this in terms of potential implementation impact
with this point in mind.  Here's some interesting observations:

If you have an update group (consistent export policy) where this is
applied, you would need to insert the timestamp on a per-peer basis.  In
order to avoid BGP path churn once you've generated an outbound timestamp,
you'd want it to stay the same to avoid churning path attributes. 

You'd thus timestamp once per outbound peer.  
You'd need to store that timestamp rather than generating it dynamically
each time you send that update.

And interestingly, the best many implementations can do is simply add the
timestamp at the time things are received or sent on the TCP session.  There
is no guarantee how closely that aligns to where it may be in TCP buffers.

Thus, tightly bound timestamps aren't very helpful here.  For the BMWG work,
if you recall, we suggested external measurement to bypass some of these
artifacts.

-- Jeff