Re: Co-Array Fortran

Gary Scott wrote:
Dan Nagle wrote:
Hello,

gary.l.scott@xxxxxxxx wrote:

Hi, I don't quite understand how the communication link with other
processes (possibly across a network) is made. It seems there should
be some sort of requirement to explicitly address network resources to
ensure the correct process is being addressed. What have I missed?


Fortran doesn't tell the compiler how to implement the standard.
It merely says what a standard-conforming program is,
and what effects the compiler must produce. How the compiler does that,
that includes, in this case, network addresses (or thread numbers)
is the compiler's job.

The standard also doesn't say how arrays are implemented.
It merely says how to declare one, and how to use it.

If you can see the funny little man behind the curtain,
the compiler's broken.

But isn't it important to at least know whether a process is "local" or "remote"? That could impact how you structure the problem solution.

Yes. Coarrays are one more kind of memory hierarchy. It's sort of
like managing cache or doing out-of core solutions (for the old
timers who remember what a core was). Personally, I think users will
have to be aware of the between image transfer times and chunk their
problem into big enough blocks to cover the transfer delays.

Coarrays seem wonderful for processors that share the same memory.
For truly remote images (like in a different city) I've heard
people talk about using coarrays for the local images and openMP
(or something similar) to handle the long distance communication.
(Or maybe it was the other way around.)

*** Hendrickson
.