Hi Chapel Team,
Sorry for the delay in sending this mail.
I would like to give more clarity on the feature that I would like to
add to the 'forall' loop.
Please find the code below implemented in Chapel using the Barrier
module available in Chapel. It was intended to use inside 'coforall',
but I am using it inside 'forall' for the sake of explanation.
use Barrier;
here.maxTaskPar = 4;
// here.maxTaskPar = 2;
const numTasks = here.maxTaskPar;
var b = new Barrier(numTasks);
var A: [1..5] int, D: [1..4] int;
var B = [1, 2, 3, 4];
proc workload (i: int) {
A[i] = B[i] - 1; // S1
b.barrier();
D[i] = A[i+1] * 2; // S2
}
forall i in 1..4 {
workload(i);
}
writeln(D);
delete b;
Assume that the barrier (b.barrier()) inside 'workload' function is
placed by the programmer so that all iterations of the 'forall' loop
finish executing S1 and then S2 (we may denote it as S1 ; S2). So the
expected values in arrays A and D are [0 1 2 3 0] and [2 4 6 0],
respectively.
Here since 'here.maxTaskPar' is set to 4 which is equal to the number
of iterations of the 'forall' loop, we get expected results. But if we
set 'here.maxTaskPar = 2', we get a different output. The code may
even hang if there are 3 tasks. The user should get a consistent
output irrespective of the number of tasks. From the users perspective
it may be easy to think that there are as many tasks as number of
iterations the 'forall' loop. But creating as many number of tasks as
iterations may be very expensive.
Here the objective is "the user should get the output as if there are
as many number of tasks as number of iterations of the 'forall' loop.
In case of Chapel we may call this kind of programming model as
'Unique Task model' unlike 'Unique Worker model' in OpenMP'.
We may solve this issue in the above example by just inlining the
'workload' function call followed by a loop distribution of the
'forall' loop across the barrier.
But the issue is more complicated when the barrier is deep inside
'if', while or 'procedures'. For instance the user may write the code
as follows.
forall i in 1..4 {
if (i % 2 == 0) {
S1;
b.barrier(); // b1
S2;
}
if (i % 2 == 1) {
S3;
b.barrier(); // b2
S4;
}
}
In 'Unique Task model' the user intends here to execute S1 and S3 in
parallel in sequence with S2 and S4 executed in parallel (denoted as
S1 || S3; S2 || S4). Even if the two barriers (b1 and b2) are placed
textually apart in the code the user wants it to act as a global
barrier which synchronizes among all the iterations of the 'forall'
loop as if there are are as many number of tasks as iterations.
Arguably this style of writing the code is useful since the programmer
can first think about the statements to be put inside various
conditions (S1 and S2 in 'if' statement and S3 and S4 in the second)
and later he/she can think about placing barriers among these
statements to satisfy the dependencies when they are executed in
parallel by the 'forall' loop.
Please share your views on such a programming model and correct me if
I made any wrong assumptions in my explanation.
If you find any interesting use cases in Chapel please share it.
Regards,
On Sat, Mar 5, 2016 at 6:15 PM, raghesh <[email protected]> wrote:
> Hi Chapel Team,
>
> I am a Post Graduate student at Indian Institute of Technology, Madras (IIT
> Madras). I would like to propose a project for Google Summer of Code this
> year. I was a GSOC student during its 2011 version and contributed to LLVM
> Polly [1].
>
> My focus is on improving the productivity of parallel programs (especially
> OpenMP programs). In that regard we have published a paper titled "Unique
> Worker model for OpenMP" [2], in ICS 2015, which facilitated placement of
> barriers (#pragma omp barrier) within parallel-for-loops (#pragma omp for).
>
> Now as GSOC project, I would like to introduce such a feature for the
> 'forall' loops in Chapel, which is more or less similar to
> parallel-for-loops in OpenMP.
>
> I am reasonably new to Chapel and started learning about the language
> features.
>
> Please share your valuable thoughts on the project idea and any comments on
> this would be of great help.
>
>
> [1] http://polly.llvm.org/
> [2] http://dl.acm.org/citation.cfm?id=2751238
>
> Regards,
> --
> Raghesh Aloor
> PhD Scholar
> PACE Lab, Dept. of CSE, IIT Madras
> http://www.cse.iitm.ac.in/~raghesh/
--
Raghesh Aloor
PhD Scholar
PACE Lab, Dept. of CSE, IIT Madras
http://www.cse.iitm.ac.in/~raghesh/
------------------------------------------------------------------------------
Transform Data into Opportunity.
Accelerate data analysis in your applications with
Intel Data Analytics Acceleration Library.
Click to learn more.
http://pubads.g.doubleclick.net/gampad/clk?id=278785111&iu=/4140
_______________________________________________
Chapel-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/chapel-users
