Re: Is parallelism for real?


In article <46940ded$1@xxxxxxxxxxxxx>, lindahl@xxxxxxx (Greg Lindahl) writes:
|> In article <f711i0$jir$1@xxxxxxxxxxxxxxxxxxxxxxx>,
|> Scott J. McCaughrin <sjmccaug@xxxxxxxxxxxxxxx> wrote:
|>
|> > Frankly, the hype about UPC and Open64
|> > fail to sway me, and UPC is equally vague in support of its claims.
|>
|> I'm confused by what you're confused by. People already spend billions
|> of dollars per year on parallel machines using the MPI and OpenMP
|> parallel programming models. UPC is just a next generation programming
|> model. Its claims are pretty obvious; folks who are familiar with the
|> existing, widely-used models find it easy to understand what's going
|> on, and if you need a tutorial on parallel programming itself, reading
|> up about UPC probably isn't the right place to start.

While you are correct about most of that, your description of UPC is
wrong and Scott McCaughrin is, regrettably, right.

I did a fair amount of investigation, and eventually cornered one of
its people at Supercomputing. As I had deduced, it is an experimental
system with essentially undefined semantics. In particular, it has
undefined semantics if there is any aliasing, but does not define what
is meant by that.

It leaves all of that to C, but C's abstract model relies heavily on
there being a single, serial CPU and is seriously ambiguous and often
inconsistent even in the serial case. C++ is much better than C90,
but C99 is worse even than C90. The new C++ should be adequate for
writing reasonably reliable parallel programs, but we are working on
it as I post.

OpenMP Fortran is pretty well the only widespread, reasonably well
defined, close-coupled shared-memory paradigm at present, with OpenMP
C usable if you have a pretty good knowledge of what not to do and
how C is implemented. There are other candidates, but none have much
market acceptance at present.

MPI called from Fortran, even C/C++ and most other languages it can be
called from is pretty solid, of course.


Regards,
Nick Maclaren.
.

Relevant Pages

  • Re: Is parallelism for real?
    ... no trouble providing code to demonstrate it. ... Greg Lindahl wrote: ... and UPC is equally vague in support of its claims. ... on, and if you need a tutorial on parallel programming itself, reading ...
    (comp.arch)
  • Re: Is parallelism for real?
    ... and UPC is equally vague in support of its claims. ... on, and if you need a tutorial on parallel programming itself, reading ...
    (comp.arch)