[DGD] DGD/MP 1.0

[Felix A. Croes]

"Christopher Allen" <ChristopherA at skotos.net> wrote:

> I was trying to find a description of the approach that you've taken with
> DGD/MP to give someone else. My less then CompSci interpretation of what
> makes it unique is that programatically you program largely as if you are
> single-threaded, but that DGD/MP will handle the issues of multiprocessor
> message passing bottlenecks and throttling through both intelligent design
> and through rollback of functions marked atomic. Obviously some care has to
> be taken with "single-thread" style code, but that largely it is much easier
> to program.
>
> However, my non-CompSci description of the merits of your approach is not
> suitable to passing on to others. Can you write up something brief about
> your unique approach and why it is superior to some of the other methods of
> programming for MP?

The compsci version:

    Maurice Herlihy: "A methodology for implementing highly concurrent
    data objects," ACM Transactions on Programming Languages and Systems,
    15(5):745-770, 1993.

The intermediate version:

    The server concurrently runs threads of finite duration on a copy
    of the data.  Once a thread completes, the changes it made in the
    copy will be committed to the original data if no other thread has
    modified that part of the original data in the meanwhile; otherwise
    the thread will be rescheduled.

Atomic functions are not involved in this, though they are conceptually
similar.

I actually started using the term "execution round" instead of "thread"
because it makes more sense, but that may be considered incomprehensible
jargon, you be the judge.  It may require explaining the LPC execution
model. :)

The non-compsci version:

    A single-threaded server is simulated using a MP server, and there
    is nothing to worry about.


In all cases, the advantage of this approach is that the great complexity
of MP design is hidden from the LPC programmer.

Regards,
Dworkin