Re: Distribution vs centralization (Was Re: Announcement: GDHE )3.2

Herman Bruyninckx <[email protected]> writes:

> On Fri, 27 Jun 2003, Anthony Mallet wrote:
> 
> > Herman Bruyninckx writes:
> > 
> > [distributed applications]
> > | I think that, almost by definition, there must be a level of control
> > | that centralizes information. Otherwise your fleet of robot can never
> > | perform a global task, or evaluate how well it does that task...
> > 
> > Just my 2c on this particular point:
> > 
> > I think that a challenge for distributed application is precisely to
> > avoid any centralized components... or it's not distributed anymore (at
> > least not fully distributed).
> 
> I think the issue is to see _how much_ you can distribute, not whether
> or not you need centralization. Of course you need centralization, or
> it's impossible for the human programmer/user of the distributed
> robotics system to keep the whole thing under control!
> The bandwidth and timing of the centralization can be quite loose, of
> course. The better the autonomy of the distributed agents, the looser
> one can centralize. 

Looking at other areas where these issues turn up (distributed lock
managers, clustering, railways, ...) the limiting case appears to be:

you need a thing that identifies to the decentralised components that
they are entitled to do whatever they want to do in the absence of
opposition.

For example, in a VMS cluster, one hard drive is this thing: if the
cluster is "broken", the side with this hard drive carries on, and the
side without it begins recovery.

In single-track-working on a railway, there is a physical token that
must be carried to enter the track section.

For a lock manager, I think it is the absence of the ownership of any of
the relevant locks by anyone else...

In the case of a robot system, the "most important" point of the system
is the flight recorder. Any component without a connection to the flight
recorder *must* enter failure/recovery mode.

Note that this is completely different from  e.g. MOSIX where a running
process has a single node it started from, which must remain in the
cluster. 


> > I'm not sure the example is relevant, but if you think about you and me
> > performing a task, there won't be any physical object that will centralize
> > information. We are just a fully distributed system that can exchange and
> > share information, agree or disagree on the way to do things, etc...
> > The information is not centralized anyhow.

Except by your CVS repository, bank account, or other transactional
storage system.

> > 
> This information is not, but the information about what task we want
> to achieve, and how well we do _is_ centralized. If it is not, how can
> we ever cooperate productively...?
> 
> Anyway, the 'agent based' infrastructure in the project proposal will
> have some level of centralization, because of the feedback component.

Are you maintaining a "current hack" version of this document on a website? 

cheers, Rich.

-- 
rich walker | technical person | Shadow Robot Company | [email protected]
front-of-tshirt space to let     251 Liverpool Road   |
                                 London  N1 1LX       | +UK 20 7700 2487