Re: [RRG] Why delaying initial packets matters

[Scott Brim <swb@employees.org>]

Please pardon my lateness ...

On 2/19/08 11:11 AM, Noel Chiappa allegedly wrote:
> As a more general comment, I don't think we're going to be able to
> predict accurately the performance of this kind of system (especially
> if it makes heavy use of caches, as I think any well-designed system
> of the kind ought to), nor the actual problems we are likely to see
> in service.
>
> Networking work has a long history of theoretical analyses that
> purport to show that approach X won't work - only to find that it
> does, in practise. (I am reminded of one series of analyses by a
> faculty member at MIT-LCS which 'showed' that contention networks
> such as Ethernet wouldn't work 'well' in high-load situations, which
> were a factor in the decision to work on token rings at MIT.
> Ethernet, as we all know, does work quite well in practise, because
> they basically never operate for long in those high-load regions.)
>
> So I think probably the best thing to do is prototype a system, and
> try it and see what happens.

Thanks for that.  The "delay" question has been hard because useful 
statistics on connection behavior are rare or confidential and we never 
know what future scenarios are going to be anyway.  So it's been hard to 
say anything useful.  But the IETF is coming, so in the meantime here 
are a few thoughts:

  - Most traffic is within a region.

  - EIDs will still be allocated by RIRs (or the equivalent) out of
    assigned chunks, so there will be a great deal of contiguousness for
    EIDs within a region.

  - ALT routers for a particular EID prefix will be fate-shared,
    situated in the region where it is predominantly found.

  - Data paths follow a topological hierarchy and the delay issue is one
    of relative differences.

So one can conjecture that for a particular session, *if* there is delay 
(not yet granted), the likely case is that the session will see far less 
than the alleged "very long paths".  There will be some outliers due to 
small chunks of EID prefix moving to a completely different region -- I 
don't think there will be many of those.  There will obviously be 
sessions between regions.  For those, at a granularity of a domain, the 
ALT control path may in fact be the same as, or at least not differ 
much, from the data path.  The relative difference in round-trip time 
will be small.

That's if there is a delay.  The only time the question arises is the 
first time a site attempts to reach a particular EID prefix, or if it is 
contacting so many prefixes that it has to do some cache management.  We 
can predict connection/cache behavior for very large and very small 
sites.  If anyone has pointers to data on connection behavior over time 
for intermediate-size sites, please say so.

This is all extrapolation from the current reality, but at least it 
tries to be based in how networks work.  We should try to get the delay 
discussion on a solid foundation of fact if possible.  Others will be 
better at that than I.

Scott

--
to unsubscribe send a message to rrg-request@psg.com with the
word 'unsubscribe' in a single line as the message text body.
archive: <http://psg.com/lists/rrg/> & ftp://psg.com/pub/lists/rrg