[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [moca] Locations?
I have more questions than answers, but maybe its a good time to put
them...
What distinguishes a distributed system from a non-distributed one is that
you can have parts of a computation failing (in various ways). Calculi can
only claim to be 'distributed' in my view if they allow you to model at
least some form of location failures.
Supposing distribution is just about failure and that a location
operator doesn't add anything to Pi. Would this make the study of
naming in distribution unimportant? Or would it add content to the
subject? The following questions rise naturally:
1- What fails? Physical resources (a computer, a component, or a whole
network)? Or conceptual entities (an activity, an event, an enterprise)?
Maybe the distinction only makes sense when they coexist, that is, when
we have two transversal structures of objects belonging to fixed
locations, but where objects of one are not associated to a fixed
location in the other. In this case, in which way could they interact?
2- What is the communication point between the world and the site that
might fail?
Anyway, I would expect a clean correspondence between those
communication points and the naming facilities we have in our hand.
My experience with distributed calculi is short and restricted to
lsd-pi, but it seems safe to say that if we're dealing with physical
resources, then there should be a tight link between named channels and
a fixed site. This is why lexical scoping seems a good tool for naming
physically distributed channels.
Does treatment of conceptual entities demand a different naming
mechanism? OK, global entities (like in DPi) might be a good tool for
naming channels as communication points to conceptual entities that
spread out over distinct physical sites. But then why would one impose
a type system that restricts channels to be associated to one site only?
The choice or design of an infrastructure should address the problems
of the world you are in, in particular the model of failures, kind of
network, etc. It is not realistic nor practical to abstract away from
these problems (as Martin and others have pointed out).
Aren't the problems of failure, synchronization, security, resource
allocation "kind of" orthogonal to each other? Isn't it reasonable to
study their properties independently, thinking modularly, and them
reassemble them into more complete systems "'a la carte"? This was our
aim with lsd-pi: concentrate on the study of resource allocation problem
in distribution.
Ana
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The "models for mobility" mailing list mailto:moca@xxxxxxxxxxxxxxx
http://www-sop.inria.fr/mimosa/personnel/Davide.Sangiorgi/moca.html