8002824276

7

lncreasing data locality ofparallel programs executed in embedded systems

Livermore loop Kernel 1 (hydro fragment) Integer loop=10000, ARRAYSIZE = sizeof(int)’1048576
i ^908 80 I 70 o 60 j| 50 = 40 .2 30 i^20 IH ^10 ' 0
			42
	SecjjenSal Execution Time	Parallel ]	Parallel SFS	^aratlel SFSwithArrai Contraction

(a)

Matrix Multiplication Integer N=2048, B=32
| 180	*181
8
■g 120 S 100 c ^80o 60 3 40 S 20 ^ 0
			60
	Sequential Execution Time	Parallel SFS	Parallel SFS with Bloeking	Parallel SFSwith Blocking SArray Conlracłion

(b)

Figurę 1. Execution time of (a) Livermore Loops Kernel 1 (b) matrix multiplication

#define SIZE 100000000	#include <omp.h>
int main(void) {	#define SIZE 100000000 int main(void)
double *a = new double[SIZE];	{
for(long i=0; i<SIZE; i++)	omp set num threads(2);
a[i]=(double)i+1;	dotible *a=new double [SIZE] ;
for(long i=0; i<(SIZE-2); i++)	for(long i=0; i<SIZE; i++)
a[i+2]=sin((a[i]*SIZE+1)/SIZE);	a[i]=(double)i+1;
return 0;	#pragma omp parallel for
}	private(j,i) shared(a)
(a) Seąuential codę	for(long j=0; j < 2 ; j++) for(long i=j; i<(SIZE-2); i+=2) a[i+2]=sin((a[i]*SIZE+1)/SIZE); return 0; } (b) Parallel codę

Figurę 2. Decreased performance in the parallel version of a simple codę