arrow-structure syntax and semantics
zccap74
zccap74@ucl.ac.uk
Mon, 7 Dec 1998 14:13:27 +0000 (GMT)
Just collecting my thoughts.....
Brian:
> -'data-flow graph' could be better explained as 'symbol flow graph'
>(or structure flow graph). the point is that the arrow-structure in
>question migrates to the location where it is used, instead of having a
>representation. maybe this concept won't work.
I think it will, you can draw analogies to thread systems:
>Such distributed threads most naturally lead to passive object style
programming models, as in Alpha; but they can be effectively employed in
>procedural models also.
>
>A distributed thread is the schedulable execution entity. It is a sequential
flow of execution through code in passive objects. It has
>
> a unique identity
> attributes -- e.g., relating to such properties as
> timeliness
> importance
> synchronization.
>
>A distributed thread's single point of execution, identity, and attributes
are maintained across address space and node boundaries, even in the
>presence of failures.
>
>A distributed thread is independent of any specific object. Distributed
threads execute in asynchronous concurrency with respect to each other,
>except as controlled by synchronizers in objects.
>
>A distributed thread extends itself outward from its root -- the point at
which it was created -- an additional segment in length, by moving its
>head -- its current execution point -- via invocation of a method on the
target object. A distributed thread retracts itself one segment inward
>toward its root by returning from an invocation.
Now admittedly your arrows are structure pointers as opposed to execution
entities (currently?) but moving the arrow structure adds a layer of
abstraction (more arrow pointers) while at the same time decreasing the layers
by associating with a location, not a representation ie. it is just changing
context, so it can work, and a lot of work has been done on migrating
threads...
It seems like the arrows representation is a good start (it solves a lot of
problems eg. type constraints etc.), however there are some things I am
worried about: you have not talked about methods of sync. (for
distriution/parallelism, no reference has been made to gcc (needs to be
considered along with semantics content in factoring out) or cooperative
scheduling.
Fare.
> What I mean is: there are lots of things that we don't have to reinvent,
> and that we shouldn't reinvent, because it's already there.
> I certainly do not mean to lessen any one's talent.
> But I'd rather learn known things from a book, and discover new things,
> than discover known things the hard way, and have no time for new things.
And a good page I have seen is:
http://www.realtime-os.com/progmod/progmod.html
(actually the whole of the site) It talks about the Alpha kernel and threading
through passive objects, as well as distribution.
>Alpha's most significant contributions to the state of the art -- both per se
and in terms of their influence on subsequent (albeit belated) OS kernel
>research -- included:
>
> a computational model specifically for time-constrained close
cooperation among arbitrarily (e.g., peer) related nodes via control flow
> based on a "distributed thread" abstraction which may span not
just (passive) objects but also address spaces and thus nodes
> with location-independent method invocation semantics -- e.g.,
each distributed thread retains its identity and attributes (including
> end-to-end timeliness constraints and parameters) -- and
syntax, as it extends and retracts itself among objects
> an efficient implementation of distributed threads using a native
kernel RPC facility (i.e., not layered on a message-passing IPC
> abstraction), which included
> asynchronous exception handling (e.g., time constraint
expirations)
> orphan detection and elimination
> dynamic, adaptive, thread resource management
> the first implementation of Jensen's
> utility function (nee' "time-value function") scheduling
framework, which provided for arbitrary mixtures of hard, general (soft),
> and no completion time constraints -- even for an individual
thread
> "best-effort" class of scheduling policies for
application-specified graceful adaptation to dynamic uncertainties such as
overloads
>
> and the provision for arbitrary application-specific real-time and
non-real-time scheduling policies, by employing a common set of kernel
> mechanisms
More when I have time...
Alexis