OpenMP Hands On

To use OpenMP in your own programs, you need a computer with more than one CPU, or a multi-core CPU and an OpenMP-capable compiler. GNU compilers later than version 4.2 support OpenMP. Also, the Sun compiler for Linux is free [2], and the Intel Compiler is free for non-commercial use [3].

Listing 5 shows an OpenMP version of the classic Hello World program. To enable OpenMP, set -fopenmp when launching GCC. Listing 8 shows the commands for building the program along with the output.

01 /* helloworld.c (OpenMP Version) */
02 #
03 #ifdef _OPENMP
04 #include <omp.h>
05 #endif
06 #include <stdio.h>
07 int main(void)
08 {
09   int i;
10 #pragma omp parallel for
11   for (i = 0; i < 4; ++i)
12   {
13     int id = omp_get_thread_num();
14     printf("Hello, World from thread %d\n", id);
15     if (id==0)
16       printf("There are %d threads\n", omp_get_num_threads());
17   }
18   return 0;
19 }

$ gcc -Wall -fopenmp helloworld.c
$ export OMP_NUM_THREADS=4
[...]
$ ./a.out
Hello World from thread 3
Hello World from thread 0
Hello World from thread 1
Hello World from thread 2
There are 4 threads

If you are using the Sun compiler, the compiler option is -xopenmp. With the Intel compiler, the option is -openmp. The Intel compiler even notifies the programmer if something has been parallelized (Listing 9).

01 $ icc -openmp helloworld.c
02 helloworld.c(8): (col. 1) remark:
03 OpenMP DEFINED LOOP WAS PARALLELIZED.

Benefits?

For an example of a performance boost with OpenMP, I'll look at a test that calculates pi [4] with the use of Gregory Leibniz's formula (Listing 7 and Figure 5). This method is by no means the most efficient for calculating pi; however, the goal here is not to be efficient but to get the CPUs to work hard.

01 /* pi-openmp.c (OpenMP version) */
02 #
03 #include <stdio.h>
04 #define STEPCOUNTER 1000000000
05 int main(int argc, char *argv[])
06   {
07   long i;
08   double pi = 0;
09   #pragma omp parallel for reduction(+: pi)
10   for (i = 0; i < STEPCOUNTER; i++) {
11     /* pi/4 = 1/1 - 1/3 + 1/5 - 1/7 + ...
12        To avoid the need to continually change
13        the sign (s=1; in each step s=s*-1),
14        we add two elements at the same time. */
15        pi += 1.0/(i*4.0 + 1.0);
16        pi -= 1.0/(i*4.0 + 3.0);
17   }
18   pi = pi * 4.0;
19   printf("Pi = %lf\n", pi);
20   return 0;
21   }

Parallelizing the for() loop with OpenMP does optimize performance (Listing 6). The program runs twice as fast with two CPUs than with one, in that more or less the whole calculation can be parallelized.

$ gcc -Wall -fopenmp -o pi-openmp pi-openmp.c
$ export OMP_NUM_THREADS=1 ; time ./pi-openmp
Pi = 3.141593
real    0m31.435s
user    0m31.430s
sys     0m0.004s
$ export OMP_NUM_THREADS=2 ; time ./pi-openmp
Pi = 3.141593
real    0m15.792s
user    0m31.414s
sys     0m0.012s

If you monitor the program with the top tool, you will see that the two CPUs really are working hard and that the pi-openmp program really does use 200 percent CPU power.

This effect will not be quite as pronounced for some problems, in which case, you might need to resort to serial execution for a large proportion of the program. Of course, your two CPUs will not be a big help in such a case, and the performance boost will be less significant. Amdahl's Law [5] (see the "Amdahl's Law" box for an explanation) applies here.

Figure 6: Amdahl's Law.