Nested Parallelism

Nested Parallelism

The OpenMP API also supports nested parallelism. Here, a parallel region is encountered inside another parallel region. This can be a useful way of attaining increased parallelism by having tasks or parallel sections that provide course-grained parallelism and then using a parallel for to gain further parallelism within the task or region. Nested parallelism is also useful in situations where the algorithm is recursive in nature.

Listing 7.37 shows an example of nested parallelism where two parallel sections con-tain parallel for directives that initialize two arrays of values. Support for nested par-allelism can be enabled through the environment variable OMP_NESTED or through a call to the routine omp_set_nested() with a value other than zero. However, support of omp_set_nested() is optional. It can be determined whether nested parallelism is enabled or not through a call to omp_get_nested().

Listing 7.37   Using Nested Parallelism to Perform Array Set Up in Parallel

#include <stdlib.h> #include <omp.h>

int main()

{

double * array1, *array2;

omp_set_nested( 1 );

#pragma omp parallel sections shared( array1, array2 )

{

#pragma omp section

{

array1 = (double*)malloc( sizeof(double)*1024*1024 );

#pragma omp parallel for shared(array1)

for ( int i=0; i<1024*1024; i++ )

{

array1[i] = i;

}

}

#pragma omp section

{

array2 = (double*)malloc( sizeof(double)*1024*512 );

#pragma omp parallel for shared(array2)

for ( int i=0; i<1024*512; i++ )

{

array2[i] = i;

}

}

}

}

However, nested parallelism is complex, so detailed discussion is left to specialist texts on the topic of OpenMP.