For a simple try at parallelization on my own outside of school, I've created a number factors calculator. I hope to eventually come up with something more creative.

Since I don't have access to parallel computers at this time, I'm using OpenMP provided by my compiler (gcc 4.8.1) and running it on my laptop (Intel Core i3-2330M). I'm using a maximum of four threads, which was determined from a call to omp_get_max_threads().

I've conducted four runs, each with four billion values and from one to four threads:

#include <cstdint>
#include <cstdlib>
#include <ctime>
#include <iomanip>
#include <iostream>
#include <map>
#include <omp.h>

void displayCompTime(std::clock_t start, std::clock_t end, std::int64_t integer, int threads)
    double elapsed = static_cast<double>(end - start) / CLOCKS_PER_SEC;

    std::cout << integer << " values and " << threads << " thread(s): "
              << std::setprecision(4) << std::fixed << elapsed << "s\n";

void calcFactors(std::int64_t integer, int threads)
    std::map<std::int64_t, std::int64_t> factors;
    std::int64_t i;

    #pragma omp parallel for num_threads(threads) default(none) \
        shared(factors, integer), private(i)
    for (i = 2; i <= integer; i++)
        if (integer % i == 0)
            factors[i] = integer / i;

int main()
    const std::int64_t integer = 4000000000;
    const int runs = 4;

    for (int i = 0; i < runs; i++)
        std::clock_t start = std::clock();
        int threads = i + 1;
        calcFactors(integer, threads);
        std::clock_t end = std::clock();
        displayCompTime(start, end, integer, threads);


4000000000 values and 1 thread(s): 67.7330s
4000000000 values and 2 thread(s): 40.7640s
4000000000 values and 3 thread(s): 32.5630s
4000000000 values and 4 thread(s): 29.7640s

Based on these results, this code doesn't appear to scale very well. I don't know if using a non-default static schedule would give faster times, and anything else would just incur additional overhead. Fortunately, I didn't need to include atomic or critical.

Would avoiding a lot of division help? I didn't try for anything else yet as this is only a start. I also wanted to see how well my laptop could handle parallelization.

Other than performance, I'm okay with any general OpenMP advice. I was sure to use some good practices for that, such as default(none) for explicitly listing the variables.


2 Answers 2


I see that OpenMP's rules will make things a little difficult here. For instance, I won't be able to concisely set i to either 2 or 3, depending on integer's parity. This is because OpenMP requires the loop counter to be set within the loop, though it can still be declared beforehand. I would otherwise have to put a ternary within the loop statement, which would look ugly. It could save one iteration, which may not make a huge difference.

As such, I might as well initialize i within the loop statement and then remove the private part from the preprocessor directive:

    shared(factors, integer)
for (std::int64_t i = 2; i <= integer; i++)

Performance-wise, I was able to get a small boost by initializing factors outside of the timed section and passing it to calcFactors(). Regardless of the time it usually takes to initialize an std::map, my runtime will always be limited by that as it's part of the serial code.

With the same division and modulus operations, I would likely not get any significant performance boost. Regardless of the thread count, the number of modulus and division operations still vary, which may explain the poor scalability over time. I also cannot test with higher thread counts due to my machine's limited number of available threads.


I have now ran this on a different machine which supports a lot more threads. Apart from some minor changes, I have utilized std::lldiv to do the division and modulo operations in one step instead of (possibly) two. To easily run it with more threads, especially with a divisible number of them, I have added commandline options to take this value and perform only one run.

for (i = 2; i <= integer; i++)
    std::lldiv_t result = std::lldiv(integer, i);

    if (result.rem == 0)
        factors[i] = result.rem;

These are my new runtimes:

4000000000 values and 1 thread(s): 63.1260s
4000000000 values and 2 thread(s): 32.0790s
4000000000 values and 4 thread(s): 15.9060s
4000000000 values and 8 thread(s): 8.3521s
4000000000 values and 16 thread(s): 6.0289s
4000000000 values and 32 thread(s): 6.1347s

The serial runtime still appears to be slow because of the division, but it's at least slightly better than the original. However, the next three runtimes appear to be quite efficient with respect to the thread counts. After that, runtimes only improve a little, but are no longer efficient. In addition, branch prediction may also be increasing runtimes, but I don't know how to get around that.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.