Re: Experiments in multi-placed root servers

[bmanning@ISI.EDU]

> 
> Bill Manning writes:
> > 	Testing is a good idea.  Do you think that doing this type of testing
> > 	is reasonable with a live root server?
> 
> I think whatever servers you test on have to be able to respond as if
> they were a root.  To get reasonable tests, you also need a
> substantial amount of pounding from a variety of places; one company
> isn't going to be able to provide that with cooperation.  As I said,
> we're looking for help in that; in other contexts, we've already done
> replay tests from log files on a variety of systems, but that doesn't
> begin to cover the full range of real-world experience.

	So, which root server is going to get permission to do this
	or are there other tactics that you will take?

> > 	You move the servers to better places.  There is no
> > 	assurence that any given technology/topology is the
> > 	end-all for Internet services.
> 
> The current system already presumes that it is better to have this
> information in many different places.  Server replication provides a
> way to increase that diversity without increasing the number of
> recognized roots.  I would certainly agree that there is no
> guarantee that it, any more than any other technology and topology,
> would answer all needs for all times.  It does, however, present
> certain advantages to the needs of our time, by providing a method
> that doesn't require we (forgive me) uproot any of the current
> servers.

	It's not clear that movment of servers to proper places
	in the topology is not the right thing to do. And I remain
	unconvinced that the method you propose will be in the
	best interests of the users of Internet services. The 
	likelyhood of the data becoming unsyncronized between
	servers is very high.

> > 	How do you dismantle this service when it no longer fits 
> > 	Internet topology or technology?  Moving a single platform
> > 	is -much- easier than killing off a replicated system. 
> 
> That actually hasn't been my experience.  If you need to move a
> replicated system, you grow it to the new space and let the old spaces
> die when they are no longer needed.  It can, of course, be harder to
> determine when a mesh of spaces is no longer needed than when a single
> space is no longer needed.

	There in lies one more problem with this approach.  Letting the
	old spaces die "when they are no longer needed" should occur 
	quickly and for all old spaces. It should not drag on over 
	indeterminant periods of time. Convergance is a good thing.

	
> > 	Current models are just that.  Remember that this was a BOF
> > 	discussing root server ops.  Of course root server placement
> > 	is the perview of IANA and not such an IETF venue. 
> 
> I think we still are discussing root server ops.  I'm interested in
> making sure that the technology works, and I believe that this group
> is an appropriate place to discuss how to to make sure of that.  If
> you feel that the overlap between this group and the root server
> advisory council isn't sufficient to cover the discussion, I'd be
> happy to forward the emails on to Jun Murai for distribution (or you
> are welcome to do so yourself).

	Jun is on this list and can make up his own mind as to your
	proposal. Root server operations and placement of root servers
	are different things.  Your proposal, to replicate large numbers
	of machines in many places, sounds like a placement issue. 
	The operation of the server is viewed in a secondary light.

	
> 			regards,
> 				Ted Hardie
> 


-- 
--bill