openMP program creates multiple threads but executes on only one core

Question

I am a new user of openMP, I have a intel i7-2670QM CPU with 8 cores on a linux ubuntu 13.10 system

My program uses nested parallelism in C to create the sum of 8 threads. As I understand it, every thread should run on it's own processor, but when I run the command top on the terminal I see that my program uses only 100% of memory (800% is expected), and in the processor view, only CPU[X] uses 100% (X is random between 0 and 7) and the other CPUs are 0.1%.

When I profile my program with Intel vtune amplifier, it shows that 7 threads were runing, but 6 of them don't use the CPU at all as they were completely IDLE.

When I try another example parallel program the threads split just fine on the cores, so I think the problem is in my code:

#include <omp.h>
void recursive_function(int k)
{
    ........
    recursive_function(...);
}
int main()
{
    omp_set_nested(1);
    #pragma omp parallel for num_threads(4)
    for(i=0;i< width * height;i++)
    {
        #pragma omp critical
        {
            ......
            // 3 simple instructions
        }
        if(i!=0)
        {
            recursive_function(i);
        }
        else
        {
            int j;
            #pragma omp parallel for num_threads(4)
            for(j=i;j< width * height;j++)
            {
                recursive_function(j);
            }
        }
    }
}

execution is made with gcc and the option -fopenmp

Your processor does not have eight cores. It has four cores and eight hyper-threads (aka logical processors). — Z boson
– Z boson, Commented Oct 23, 2014 at 13:46
ok, but even though the program shoul use the whole processor, not just one logical processor — CSNooB
– CSNooB, Commented Oct 23, 2014 at 14:03
You're right. I don't know what's causing your problem based on the information you have provided. — Z boson
– Z boson, Commented Oct 23, 2014 at 19:52

Blue Granny · Accepted Answer · 2016-08-11 12:20:32Z

2

Have you tried setting GOMP_CPU_AFFINITY?

It may be the scheduler that's not working properly.

EDIT: Changed to GOMP_CPU_AFFINITY as per Haralds comment. He provides a link there as well.

edited Aug 11, 2016 at 12:20

answered Aug 10, 2016 at 6:28

Blue Granny

8002 gold badges8 silver badges25 bronze badges

Sign up to request clarification or add additional context in comments.

2 Comments

Harald Over a year ago

The KMP_AFFINITY is for the Intel OpenMP runtime, not for GNU runtime. For GNU runtime you have GOMP_CPU_AFFINITY.

Blue Granny Over a year ago

Thanks for pointing this out. I forgot about this distinction

Harald · Accepted Answer · 2016-08-10 10:18:17Z

1

Notice that you go into the nested region only for i==0.

That means that the outer loop is executed by a team of 4 threads (call it T1). Whenever a thread from T1 executes the iteration i=0 (let's call that thread TH1), TH1 will go into the else and then it will create a parallel region with a team of 4 threads (call it T2). At this moment team T1 has the 3 remaining threads executing the cases where i!=0 and team T2 has 4 threads (including TH1) that executes the innermost parallel region. That sums up to 7 threads.

With respect to threads being idle, that completely depends on the work they have to execute -- that is, recursive_function().

answered Aug 10, 2016 at 10:18

Harald

3,2062 gold badges28 silver badges35 bronze badges

Collectives™ on Stack Overflow

openMP program creates multiple threads but executes on only one core

2 Answers 2

2 Comments

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

2 Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Related