Question

Hello,

I have created a multithreaded application for multiplying two matrices using pthreads,but to my surprise the multithreaded program is taking much time than my expectation.

I dnt know where is the problem in my code,the code snippet is given below::

#include "pthreads.h"
#include "cv.h"
#include "cxcore.h"

CvMat * matA;       /* first matrix */
CvMat * matB;       /* second matrix */
CvMat * matRes;     /* result matrix */

int size_x_a; /* this variable will be used for the first  dimension */
int size_y_a; /* this variable will be used for the second dimension */

int size_x_b,size_y_b;
int size_x_res;
int size_y_res;

struct v {
  int i; /* row */
  int j; /* column */
};


void *printThreadID(void *threadid)
{
/*long id = (long) threadid;
//printf("Thread ID: %ld\n", id);

arrZ[id] = arrX[id] + arrY[id];

pthread_exit(NULL);*/
return 0;
}

int main()
{
/* assigining the values of sizes */
size_x_a = 200;
size_y_a = 200;
size_x_b = 200;
size_y_b = 200;

/* resultant matrix dimensions */
size_x_res = size_x_a;
size_y_res = size_y_b;

matA = cvCreateMat(size_x_a,size_y_a,CV_64FC1);
matB = cvCreateMat(size_x_b,size_y_b,CV_64FC1);
matRes = cvCreateMat(size_x_res,size_y_res,CV_64FC1);

pthread_t thread1;
pthread_t thread2;
pthread_t multThread[200][200];

int res1;
int res2;
int mulRes;
/*******************************************************************************/ 

/*Creating a thread*/
res1 = pthread_create(&thread1,NULL,initializeA,(void*)matA);
if(res1!=0)
{
    perror("thread creation of thread1 failed");
    exit(EXIT_FAILURE);
}


/*Creating a thread*/
res2 = pthread_create(&thread2,NULL,initializeB,(void*)matB);

if(res2!=0)
{
    perror("thread creation of thread2 failed");
    exit(EXIT_FAILURE);
}


pthread_join(thread1,NULL);
pthread_join(thread2,NULL);

/*Multiplication of matrices*/
for(int i=0;i<size_x_a;i++)
    {
  for(int j=0;j<size_y_b;j++)
      {
      struct v * data = (struct v*)malloc(sizeof(struct v));
      data->i = i;
      data->j = j;

mulRes = pthread_create(&multThread[i][j],NULL,multiplication,  (void*)data);
       }
    }

for(int i=0;i<size_x_a;i++)
{
for(int j=0;j<size_y_b;j++)
    {
    pthread_join(multThread[i][j],NULL);    
    }
}


for(int i =0;i<size_x_a;i++)
{
    for(int j = 0;j<size_y_a;j++)
    {
        printf("%f ",cvmGet(matA,i,j));
    }
}
return 0;
}

void * multiplication(void * param)
{
struct v * data = (struct v *)param;
double sum =0;
for(int k=0;k<size_x_a;k++)
    sum += cvmGet(matA,data->i,k) * cvmGet(matB,k,data->j); 

cvmSet(matRes,data->i,data->j,sum);
pthread_exit(0);

return 0;
}

void * initializeA(void * arg)
{
CvMat * matA  = (CvMat*)arg;
//matA = (CvMat*)malloc(size_x_a * sizeof(CvMat *));

/*initialiazing random values*/
for (int i = 0; i < size_x_a; i++) 
{
 for (int j = 0; j < size_y_a; j++) 
 {
    cvmSet(matA,i,j,size_y_a + j); /* just some unique number for each element */
 }
}
return 0;
}

void * initializeB(void * arg)
{
CvMat* matB  = (CvMat*)arg;
//matB = (CvMat*)malloc(size_x_b * sizeof(CvMat *));

/*initialiazing random values*/
for (int i = 0; i < size_x_b; i++) 
{
  for (int j = 0; j < size_y_b; j++) 
  {
    cvmSet(matB,i,j,size_y_b + j); /* just some unique number for each element */
  }
}
return 0;
}

void * initializeRes(void * arg)
{
CvMat * res  = (CvMat*)arg;
//res = (CvMat*)malloc(size_x_res * sizeof(CvMat *));

/* for matrix matRes, allocate storage for an array of ints */
for (int i = 0; i < size_x_res; i++) 
{
    for (int j = 0; j < size_y_res; j++) 
    {
        cvmSet(matRes,i,j,0);
    }
}
return 0;
}

I am doing this multithreading for the first time.
Kindly help me with this,any suggestion or correction will be very helpful.

Thanks in advance.

Answer 1

You’re creating ALOT of threads, which will involve lots of context switches. If each thread is doing pure calculations, and wont involve any sort of waiting (like networking, sockets, etc) there is no reason why threading will be faster than not threaded. Unless of course you are on a multi CPU/core machine, then you should create one thread per core. With this sort of processing, more threads than cores will just slow it down.

What you could do is divide the work-set into tasks that can be enqueued, and have worker threads (one/CPU core) that will pull the tasks off of a common worker queue. This is a standard producer/consumer problem.

Here is some generic info about the producer/consumer problem.

Its been a long time since Ive done matrix multiplication, so bear with me 🙂 It appears that you could divide the following into separate tasks:

/*Multiplication of matrices*/
for(int i=0;i<size_x_a;i++)
    {
  for(int j=0;j<size_y_b;j++)
      {
      struct v * data = (struct v*)malloc(sizeof(struct v));
      data->i = i;
      data->j = j;

      /* Instead of creating a thread, create a task and put it on the queue
       * mulRes = pthread_create(&multThread[i][j],NULL,multiplication,  (void*)data);
       */

      /* Im not going to implement the queue here, since there are several available
       * But remember that the queue access MUST be mutex protected. */
      enqueue_task(data);
       }
    }

Previously, you will have to have created what is called the thread-pool (the worker threads, one per CPU core), whose worker function will try to pull off the queue and execute the work. There are ways to do this with pthread conditional variables, whereby the threads are blocked/waiting on the cond var if the queue is empty, and once the queue is populated, then the cond var is signalled, thus releasing the threads so they can start working.

If this is not a logical division of work, and you cant find one, then perhaps this problem is not suitable for multi-threading.