CODEWITHSUNDEEP
cgccopenmp
user1620200
user1620200

Reputation: 103

OpenMP and C parallel for loop: why does my code slow down when using OpenMP?

I'm new here and a beginner level programmer in C. I'm having some problem with using openmp to speedup the for-loop. Below is simple example:

#include <stdlib.h>
#include <stdio.h>
#include <gsl/gsl_rng.h>
#include <omp.h>

gsl_rng *rng;

main()
{
int i, M=100000000;
double tmp;

/* initialize RNG */
gsl_rng_env_setup();
rng = gsl_rng_alloc (gsl_rng_taus);
gsl_rng_set (rng,(unsigned long int)791526599);

// option 1: parallel        
  #pragma omp parallel for default(shared) private( i, tmp ) schedule(dynamic)
  for(i=0;i<=M-1;i++){
     tmp=gsl_ran_gamma_mt(rng, 4, 1./3 );
  }


// option 2: sequential       
  for(i=0;i<=M-1;i++){
     tmp=gsl_ran_gamma_mt(rng, 4, 1./3 );
  }
}

The code draws from a gamma random distribution for M iterations. It turns out the parallel approach with openmp (option 1) takes about 1 minute while the sequential approach (option 2) takes only 20 seconds. While running with openmp, I can see the cpu usage is 800% ( the server I'm using has 8 CPUs ). And the system is linux with GCC 4.1.3. The compile command I'm using is gcc -fopenmp -lgsl -lgslcblas -lm (I'm using GSL )

Am I doing something wrong? Please help me! Thanks!

P.S. As pointed out by some users, it might be caused by rng. But even if I replace

tmp=gsl_ran_gamma_mt(rng, 4, 1./3 );

by say

tmp=1000*10000;

the problem still there...

Upvotes: 9

Views: 2574

Answers (3)

user1620200
user1620200

Reputation: 103

Again thanks everyone for helping. I just found out that if I get rid of 

schedule(dynamic)

in the code, the problem disapears. But why is that?

Upvotes: 1

Greg Inozemtsev
Greg Inozemtsev

Reputation: 4671

Your rng variable is shared, so the threads are spending all their time waiting to be able to use the random number generator. Give each thread a separate instance of the RNG. This will probably mean making the RNG initialization code run in parallel as well.

Upvotes: 5

Konrad Rudolph
Konrad Rudolph

Reputation: 546015

gsl_ran_gamma_mt probably locks on rng to prevent concurrency issues (if it didn’t, your parallel code probably contains a race condition and thus yields wrong results). The solution then would be to have a separate rng instance for each thread, thus avoiding locking.

Upvotes: 12

Related Questions