UNIX for Dummies Questions & Answers

Hello,

I run my pthread code on Linux with 4 processors. However, the speed up is only 2 times.

The code is about solving equation (G+s(i)C)z(i)=B*us(i), i=1,...,n. Here G,C are m*m matrix, B*us(i) is a m*1 vector and s(i) are n different numbers. I need to solve the equation n times to get z(1)...z(n). Here, I use multithread to solve the n equations.
eg. Now I have 4 thread and n=12. Thread(1) solves 4 equations (G+s(1,2,3,4)C)X(1,2,3,4)=B. Thread(2) solves (G+s(5,6,7,8)C)X(5,6,7,8)=B....

I use pthread_creat() and pthread_join(). However, for 4thread, it is not as 4 times faster as 1 thread, only 2 times instead. For 2 thread, it is about 1.5 times as 1 thread. What I have observed that, to solve one equation under 4 thread is much slower than solving one equation using 1 thread.

Can anybody tell me what is the reason? Thanks a lot.

Below is the code. Thanks,

typedef struct{
int thread_no;
int allthread;
mat *G,*C,*B;
mat *us, *Z;
vec *samples;
double *Control, *Info;
} parm;


void sampleLU(int thread_no, int allthread, mat *G, mat *C, mat *B, mat &us, mat &Z, vec &samples, double Control[], double Info[])
{
Real_Timer lu_symbolic_init, lu_symbolic_free, lu_numerical, lu_solve_time;
Real_Timer sCpG_run_time;
int np = samples.size();
int nDim = B->m;
int start, stop;

start = thread_no * (int)(np/allthread);
stop = start + (int)(np/allthread) - 1;
if( thread_no == allthread-1 ) stop = np-1;
for (int i = start; i<=stop; i++ ){
cs *A;
if(thread_no == 0) {
sCpG_run_time.start();
A = G+sample(i)*C;
sCpG_run_time.stop();
}
else{
A = G+sample(i)*C;
}

/* LU decomposition */
...
if(i == start){
if(thread_no == 0) {
lu_symbolic_init.start();
A = LU;
lu_symbolic_init.stop();
}
else {
A = LU
}
}

/* solve Az = b */
double* z = new double[nDim];
vec b(nDim);
b.zeros();
if(thread_no == 0) lu_solve_time.start();
LUz=b
vec zz(z, nDim);
delete [] z;
Z.set_col(i, zz);
...
}
if(thread_no == 0){
std::cout << "Thread" << thread_no << " sC+G \t:" << sCpG_run_time.get_time() << std::endl;
std::cout << "Thread" << thread_no << " symbolic initial time: \t"<<lu_symbolic_init.get_time()<<std::endl;
std::cout << "Thread" << thread_no << " LU decomposition time \t:" << lu_numerical.get_time() << std::endl;
std::cout << "Thread" << thread_no << " symbolic free time: \t"<<lu_symbolic_init.get_time()<<std::endl;
std::cout << "Thread" << thread_no << " LU solve time \t: " << lu_solve_time.get_time() <<std::endl;

}
}
void * psampleLU(void *arg)
{
parm *p = (parm *)arg;
sampleLU(p->thread_no, p->allthread, p->G, p->C, p->B, *(p->us), *(p->Z), *(p->samples), p->Control, p->Info);
return NULL;
}

You have not shown us all your code. Please show all your code and provide information about your OS version, compiler version and how you compiled your executable.

Linux quad core Intel Xeon (2.99GHz, 16G memory)
compiler gcc version 4.1.2 20071124 (Red Hat 4.1.2-42)

I only use -lpthread

Code:

typedef struct{
int thread_no;
int allthread;
mat *G,*C,*B;
mat *us, *Z;
vec *samples;
double *Control, *Info;
} parm;


void sampleLU(int thread_no, int allthread, mat *G, mat *C, mat *B, mat &us, mat &Z, vec &samples, double Control[], double Info[])
{
Real_Timer lu_symbolic_init, lu_symbolic_free, lu_numerical, lu_solve_time;
Real_Timer sCpG_run_time;
int np = samples.size();
int nDim = B->m;
int start, stop;

start = thread_no * (int)(np/allthread);
stop = start + (int)(np/allthread) - 1;
if( thread_no == allthread-1 ) stop = np-1;
for (int i = start; i<=stop; i++ ){
cs *A;
if(thread_no == 0) {
sCpG_run_time.start();
A = G+sample(i)*C;
sCpG_run_time.stop();
}
else{
A = G+sample(i)*C;
}

/* LU decomposition */
...
if(i == start){
if(thread_no == 0) {
lu_symbolic_init.start();
A = LU;
lu_symbolic_init.stop();
}
else {
A = LU
}
}

/* solve Az = b */
double* z = new double[nDim];
vec b(nDim);
b.zeros();
if(thread_no == 0) lu_solve_time.start();
LUz=b
vec zz(z, nDim);
delete [] z;
Z.set_col(i, zz);
...
}
if(thread_no == 0){
std::cout << "Thread" << thread_no << " sC+G \t:" << sCpG_run_time.get_time() << std::endl;
std::cout << "Thread" << thread_no << " symbolic initial time: \t"<<lu_symbolic_init.get_time()<<std::endl;
std::cout << "Thread" << thread_no << " LU decomposition time \t:" << lu_numerical.get_time() << std::endl;
std::cout << "Thread" << thread_no << " symbolic free time: \t"<<lu_symbolic_init.get_time()<<std::endl;
std::cout << "Thread" << thread_no << " LU solve time \t: " << lu_solve_time.get_time() <<std::endl;

}
}
void * psampleLU(void *arg)
{
parm *p = (parm *)arg;
sampleLU(p->thread_no, p->allthread, p->G, p->C, p->B, *(p->us), *(p->Z), *(p->samples), p->Control, p->Info);
return NULL;
}

int main(int argc, char* argv[]){
pthread_t *threads;
pthread_attr_t pthread_custom_attr;

parm *arg;

threads = (pthread_t *)malloc(allthread * sizeof(pthread_t));
pthread_attr_init(&pthread_custom_attr);

arg = (parm *)malloc(allthread * sizeof(parm));

for ( int i=0; i<allthread; i++){
arg[i].thread_no = i;
arg[i].allthread = allthread;
arg[i].G = G;
arg[i].C = C;
arg[i].B = B;
arg[i].us = &us;
arg[i].Z = &Z;
arg[i].samples = &samples;
arg[i].Control = Control;
arg[i].Info = Info;
/*arg[i].dim = NDIM;
arg[i].a = &a;
arg[i].b = &b;
arg[i].c = &c;*/
if(i == 0) pthread_samplingLU.start();
pthread_create(&threads[i], &pthread_custom_attr, psampleLU,(void*)(arg+i));
//pthread_create(&threads[i], NULL, psampleLU,(void*)(arg+i));
}

for ( int i=0; i<allthread; i++){
pthread_join(threads[i], NULL);
if(i == 0) pthread_samplingLU.stop();
}

free(arg);
}
}