Question

I’m having a Segmentation Fault when accessing an array inside a for loop.
What I’m trying to do is to generate all subsequences of a DNA string.

It was happening when I created the array inside the for. After reading for a while, I found out that the openmp limits the stack size, so it would be safer to use the heap instead. So I change the code to use malloc, but the problem persists.

This is the full code:

#include <limits.h>
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <omp.h>

#define DNA_SIZE 26 
#define DNA "AGTC"

static char** powerset(int argc, char* argv)
{
    unsigned int i, j, bits, i_max = 1U << argc;

    if (argc >= sizeof(i) * CHAR_BIT) {
        fprintf(stderr, "Error: set too large\n");
        exit(1);
    }
    omp_set_num_threads(2);
    char** subsequences = malloc(i_max*sizeof(char*));

    #pragma omp parallel for shared(subsequences, argv) 
    for (i = 0; i < i_max ; ++i) {
        //printf("{");
        int characters = 0;
        for (bits=i; bits ; bits>>=1)
            if (bits & 1)
                ++characters;

        //This is the line where the error is happening. 
        char *ss = malloc(characters+1 * sizeof(char)*16);//the *16 is just to save the cache lin       

        int ssindex = 0;

        for (bits = i, j=0; bits; bits >>= 1, ++j) {
            if (bits & 1) {
                //char a = argv[j];
                ss[ssindex++] = argv[j] ;
            } 
        }
        ss[ssindex] = '\0';
        subsequences[i] = ss;       
    }
    return subsequences;
}

char* getdna()
{
    int i;

    char *dna = (char *)malloc((DNA_SIZE+1) * sizeof(char));

    for(i = 0; i < DNA_SIZE; i++)
    {
        int randomDNA = rand() % 4;
        dna[i] = DNA[randomDNA];
    }

    dna[DNA_SIZE] = '\0';

    return dna;
}

void printResult(char** ss, int size)
{
    //PRINTING THE SUBSEQUENCES
    printf("SUBSEQUENCES FOUND:\r\n");
    int i;
    for(i = 0; i < size; i++)
    {
        printf("%i.\t{ %s } \r\n",i+1 , ss[i]);
        free(ss[i]);
    }
    free(ss);
}

int main(int argc, char* argv[])
{
    srand(time(NULL));
    double starttime, stoptime;
    starttime = omp_get_wtime();
    char* a = getdna();
    printf("%s\r\n", a);
    int size = pow(2, DNA_SIZE);
    printf("number of subsequences: %i\r\n", size);

    char** subsequences = powerset(DNA_SIZE, a);    
    //todo: make it optional printing to the stdout or saving to a file
    //printResult(subsequences, size);
    stoptime = omp_get_wtime();

    printf("Tempo de execucao: %3.2f segundos\n\n", stoptime-starttime);
    printf("Numero de sequencias geradas: %i\n\n", size);
    free(a);
    return 0;
}

I also tried to make the malloc line critical with the #pragma omp critical which didn’t help.
Also I tried to compile with -mstackrealign which also didn’t work.

Appreciate all the help.

Answer 1

You should use a more efficient thread-safe memory management.

Applications can use either malloc() and free() explicitly, or implicitly in the compiler-generated code for dynamic/allocatable arrays, vectorized intrinsics, and so on.

The thread-safe malloc() and free() in some libc implementations carry a high synchronization overhead caused by internal locking. Faster allocators for multi-threaded applications exist. For instance, on Solaris multithreaded applications should be linked with the "MT-hot" allocator mtmalloc, (i.e., link with -lmtmalloc to use mtmalloc instead of the default libc allocator). glibc, used on Linux and some OpenSolaris and FreeBSD distributions with GNU userlands, uses a modified ptmalloc2 allocator, which is based on Doug Lea’s dlmalloc. It uses multiple memory arenas to achieve near lock-free behavior. It can also be configured to use per-thread arenas and some distributions, notably RHEL 6 and derivates, have that feature enabled.

static char** powerset(int argc, char* argv)
{
    int i, j, bits, i_max = 1U << argc;

    if (argc >= sizeof(i) * CHAR_BIT) {
        fprintf(stderr, "Error: set too large\n");
        exit(1);
    }
    omp_set_num_threads(2);
    
    
    char** subsequences = malloc(i_max*sizeof(char*));
    
    int characters = 0;
    for (i = 0; i < i_max ; ++i)
    {
         for (bits=i; bits ; bits>>=1)
            if (bits & 1)
                ++characters;
         
        subsequences[i] = malloc(characters+1 * sizeof(char)*16);
        characters = 0;
    }
    
    
    #pragma omp parallel for shared(subsequences, argv) private(j,bits)
    for (i = 0; i < i_max; ++i)
    {     

        int ssindex = 0;

        for (bits = i, j=0; bits; bits >>= 1, ++j) {
            if (bits & 1) {
                subsequences[i][ssindex++] = argv[j] ;
            } 
        }
       subsequences[i][ssindex] = '\0';
    }
    
    return subsequences;
}

I create (and allocate) the desired data before the parallel region, and then made the remaining calculations. The version above running with 12 threads in a 24 core machine takes "Tempo de execucao: 9.44 segundos".

However, when I try to parallelize the following code:

   #pragma omp parallel for shared(subsequences) private(bits,characters)
    for (i = 0; i < i_max ; ++i)
            {
                 for (bits=i; bits ; bits>>=1)
                    if (bits & 1)
                        ++characters;
                 
                subsequences[i] = malloc(characters+1 * sizeof(char)*16);
                characters = 0;
            }

it take "Tempo de execucao: 10.19 segundos"

As you can see calling malloc in parallel leads to slower times.

Eventually, you would have had problems with the fact that each sub-malloc was trying to allocate (characters+1*DNA_SIZE*sizeof(char)) rather than ((characters+1)*DNA_SIZE*sizeof(char)), and the multiplying by a factor for cache line size is not necessary inside the parallel section if I understand what you were trying to avoid.

There also seems to be some issue with this piece of code:

for (bits = i, j=0; bits; bits >>= 1, ++j) {
    if (bits & 1) {
        //char a = argv[j];
        ss[ssindex++] = argv[j] ;
    }
}

With this code, j sometimes hits DNA_SIZE or DNA_SIZE+1, resulting in reading argv[j] going off the end of the array. (Also, using argc and argv as names for arguments in this function is somewhat confusing.)