Dear Mohammed.
You are welcome.
A simple way to understand loop chunking:
Replace:
foreach (i: [1..n]) {
S
}
==>
foreach(j: [1..2]) {
for (k = [1..n/2]) {// assuming n is even
i = (j-1)*n/2 + k
S
}
}
For your specific example - a simple way: create a vector of regions (in
this case a vector of two regions).
> for((i,j):Point in a.region)
> {
> async
> {
=>
for (region r: a.regVector) // code to create a.regVector - not shown.
async
for ((i,j ) : r) {
...
}
}
You can of course optimize the code - so that you can avoid using
collection class like Vector. By replacing the outer for loop to iterate
over 'n' iterations and the inner for loop to iterate over a region whose
dimensions are (1000/n x 1000).
Hope this helps.
Warm regards,
Krishna.
|------------>
| From: |
|------------>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
|mohammed elsaeedy <mohammed.elsae...@kaust.edu.sa>
|
>--------------------------------------------------------------------------------------------------------------------------------------------------|
|------------>
| To: |
|------------>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
|Mailing list for users of the X10 programming language
<x10-users@lists.sourceforge.net>
|
>--------------------------------------------------------------------------------------------------------------------------------------------------|
|------------>
| Date: |
|------------>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
|07/05/2010 04:36 PM
|
>--------------------------------------------------------------------------------------------------------------------------------------------------|
|------------>
| Subject: |
|------------>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
|Re: [X10-users] Async Parallelism????
|
>--------------------------------------------------------------------------------------------------------------------------------------------------|
Dear Krishna,
Thank you very much for your reply, and yes it seems very convincing,
but do you know how to modify my code
to be divided to two phases? I mean how to divide the loop part into two
parts (which are the 2 threads)
Thank you,
Regards,
Mohammed
On Mon, Jul 5, 2010 at 6:27 AM, Krishna Nandivada Venkata <
nvkris...@in.ibm.com> wrote:
> Hi.
> You are essentially creating a lot of activities, but the fact remains
you
> have only two compute threads. That is you specify your ideal parallelism
> (1000x1000 threads), but the useful parallelism is just two threads - you
> have only two hardware cores or threads. So there is a significant
overhead
> in terms of activity creation/scheduling/termination.
>
> A simple but effective means to handle it is via loop chunking (see
> http://portal.acm.org/citation.cfm?id=1542275.1542304). I don't think it
> is
> yet implemented in the X10 compiler. So for the time being you may want
to
> do it manually - replace the parallel loop which creates a large number
of
> activities with one that creates fewer number of activities.
>
> Warm regards,
> Krishna.
>
> |------------>
> | From: |
> |------------>
>
>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
> |mohammed elsaeedy <mohammed.elsae...@kaust.edu.sa>
>
|
>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
> |------------>
> | To: |
> |------------>
>
>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
> |Mailing list for users of the X10 programming language <
> x10-users@lists.sourceforge.net>
> |
>
>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
> |------------>
> | Date: |
> |------------>
>
>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
> |07/05/2010 03:49 PM
> |
>
>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
> |------------>
> | Subject: |
> |------------>
>
>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
> |Re: [X10-users] Async Parallelism????
> |
>
>
>--------------------------------------------------------------------------------------------------------------------------------------------------|
>
>
>
>
>
> I'm sorry but the previous mail was sent by mistake incomplete so heres
my
> message:
>
> Dear List,
>
>
> I have this very weird behavior going on, I'm just implementing this
> simple Matrix-Vector Multiplication program
> and I'm measuring the time consumed by my program, so heres my parallel
> code
> with asyncs :
> *
> val n:int=1000;
> val a= new Array[Int]([0..n,0..n],((i,j):Point)=>j); //Matrix a
2
> dim
> val b= new Array[Int]([0..n],((i):Point)=>i); //Vector b
> 1
> dim
> val result = new Array[Int]([0..n],((i):Point)=>0); //Vector
> result 1 dim result= ab
>
> val h = new Hello();
> val begin:Long = Timer.nanoTime();
>
> finish
> {
> for((i,j):Point in a.region)
> {
> async
> {
> val value:int;
> finish
> value= h.computeMult(a,b,i,j); //this is just
> a
> method that calculates the multiplication of each matrix row element with
> the
> //
> corresponding vector column element
> atomic result(i)+=value; // atomic
> sum
> up of the corresponding result element
> }
> }
> }
>
> val end:Long = Timer.nanoTime();*
>
>
> So as you can see the size of the matrix and the vector is 1000, the
> execution time with "async, finish, atomic" (works in one place but
> multi-activities) is *11.872344* sec where as if it worked sequentially
(by
> removing all "async", "finish", "atomic") it gives a better result which
is
> *2.39275* secs, and ofcourse as I increase the size of the dimensions it
> goes worse. I'm working on a dual core machine.
>
> How is this possible? please advise.
>
> On Mon, Jul 5, 2010 at 5:15 AM, mohammed elsaeedy <
> mohammed.elsae...@kaust.edu.sa> wrote:
>
> > Dear List,
> >
> >
> > I have this very weird behavior going on, I'm just implementing
this
> > simple Matrix-Vector Multiplication program
> > and I'm measuring the time consumed by my program, so heres my parallel
> > code with asyncs :
> > *
> > val n:int=1000;
> > // 2
> > val a= new Array[Int]([0..n,0..n],((i,j):Point)=>j);
> > val b= new Array[Int]([0..n],((i):Point)=>i);
> > val result = new Array[Int]([0..n],((i):Point)=>0);
> >
> > val h = new Hello();
> > val begin:Long = Timer.nanoTime();
> >
> > finish
> > {
> > for((i,j):Point in a.region)
> > {
> > async
> > {
> > val value:int;
> > finish
> > value= h.computeMult(a,b,i,j); //this is
> just
> > a method that calculates
> > atomic result(i)+=value;
> > }
> > }
> > }
> >
> > val end:Long = Timer.nanoTime();*
> >
> > --
> > Thank you for your concern.
> > Regards,
> > Mohammed El Sayed
> >
>
>
>
> --
> Thank you for your concern.
> Regards,
> Mohammed El Sayed
>
>
------------------------------------------------------------------------------
>
> This SF.net email is sponsored by Sprint
> What will you do first with EVO, the first 4G phone?
> Visit sprint.com/first -- http://p.sf.net/sfu/sprint-com-first
> _______________________________________________
> X10-users mailing list
> X10-users@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/x10-users
>
>
>
>
>
------------------------------------------------------------------------------
> This SF.net email is sponsored by Sprint
> What will you do first with EVO, the first 4G phone?
> Visit sprint.com/first -- http://p.sf.net/sfu/sprint-com-first
> _______________________________________________
> X10-users mailing list
> X10-users@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/x10-users
>
--
Thank you for your concern.
Regards,
Mohammed El Sayed
------------------------------------------------------------------------------
This SF.net email is sponsored by Sprint
What will you do first with EVO, the first 4G phone?
Visit sprint.com/first -- http://p.sf.net/sfu/sprint-com-first
_______________________________________________
X10-users mailing list
X10-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/x10-users
------------------------------------------------------------------------------
This SF.net email is sponsored by Sprint
What will you do first with EVO, the first 4G phone?
Visit sprint.com/first -- http://p.sf.net/sfu/sprint-com-first
_______________________________________________
X10-users mailing list
X10-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/x10-users
