PS -- The "typically" is because the author of a domain map can implement
the parallelism in the forall loop any way they want to. The standard
distributions that we supply are intended to target both inter- and
intra-locale parallelism.
On Mon, 29 Aug 2016, Brad Chamberlain wrote:
Hi Hui --
Forall loops overdistributed domains/arrays typically enable parallelism both
across locales and within a compute node. So the better sketch of how it
would be rewritten
coforall loc in Locales {
on loc {
forall msg in local_MessageSpace {
writeln();
}
}
}
Where that inner 'forall' would use multicore parallelism within the locale,
much as a forall loop over a default domain/array would.
Hope this helps,
-Brad
On Fri, 26 Aug 2016, Hui Zhang wrote:
Hello,
I have a question about forall, considering the simple example of
hello4-datapar-dist.chpl:
config const numMessages = 100;
const MessageSpace = {1..numMessages} dmapped Cyclic(startIdx=1);
forall msg in MessageSpace do
writeln("Hello, world! (from iteration ", msg, " of ", numMessages,
"owned by locale ", here.id, " of ", numLocales, ")");
Here, after the domain is dmapped, the MessageSpace will be distributed.
For example, we have 10 nodes, and 2 cores per node. Will the forall loop
ends up getting a parallelism of 20 ?
I ran it and the output is in arbitary order, so does the forall implicitly
enable both the parallelism among locales and the parallelism within a
single locale ? In other words, does it actually have similar effect as the
following style:
coforall loc in Locales {
on loc {
for msg in local_MessageSpace {
writeln();
}
}
}
*the local_MessageSpace will be determined by here.id, MessageSpace and the
domain mapping it uses(e.g Cyclic), sometimes hard to determne
Thanks !
--
Best regards
Hui Zhang
------------------------------------------------------------------------------
_______________________________________________
Chapel-developers mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/chapel-developers
