John,
This communication issue is not specific to domains/arrays, although
it is more visible with domains/arrays.
Consider this simple example:
class TestClass {
var initArr: [...] real;
proc DoSomething() {
forall ... {
... this.initArr[...] ...
}
}
}
const TC = new TestClass();
TC.DoSomething();
In order to compute "this.initArr[...]", currently we need to get
the pointer to initArr (meta-)data. That pointer is stored in 'this'
object.
On those iterations of the 'forall' that are on a different locale
than 'this', fetching that pointer from 'this' results in a remote get.
This is, I believe, what bytes you here.
Avoiding these remote fetches is a simple optimization, which we
do not do currently. As a workaround, you can pass 'initArr' explicitly
into DoSomething():
class TestClass {
var initArr: [...] real;
proc DoSomething(initArr) {
forall ... {
... initArr[...] ... // reference the formal
}
}
}
const TC = new TestClass();
TC.DoSomething(TC.initArr); // pass initArr
As an aside, since in your example initArr and resArr are
Block-distributed, they are "privatized". So what we really fetch
is their integer 'pid' fields; the rest is done locally.
If my explanation does not seem to help, please let us know.
Vass
On 03/05/15 14:18, John MacFrenz wrote:
> Hi,
>
> While writing a test case to compare my new distribution with the block
> distribution I ran into some performance problems when having distributed
> domains, arrays and domainmaps inside a class.
>
> If domain map, domains and arrays are members of a class, and computing is
> done in a method, it seems to generate much more communication than when not
> using class. Gets on the second locale are 8022 without classes and 20222
> using the class thing on my computer. --cache-remote reduces communications,
> but the ratio of gets (or get_nbs) stays same.
>
> Code demonstrating this problem is as an attachment. I ran the test with two
> locales. Notice the number of gets on the second locale when using the class
> approach.
>
> Is this just a limitation of chapel implementation, or does the class
> approach require some special tricks...?
------------------------------------------------------------------------------
Dive into the World of Parallel Programming The Go Parallel Website, sponsored
by Intel and developed in partnership with Slashdot Media, is your hub for all
things parallel software development, from weekly thought leadership blogs to
news, videos, case studies, tutorials and more. Take a look and join the
conversation now. http://goparallel.sourceforge.net/
_______________________________________________
Chapel-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/chapel-users
