Pthread vs OpenMPの性能比較してみた （科学計算）

ふと思いついてPthreadとOpenMPの性能の比較をしてみた。　blasを使った倍精度の内積計算。

AMD: C-50 (2core)
gcc: 4.7.0 (gmp-5.0.5)

real:
     serial 1.079 s
     Omp   0.830 s ( 2 thread )
     pthread 1.225 s ( 2 thread )

ネットブックのじゃ、スレッド呼出のオーバヘッドの方が大きいのか？ pthreadがなんだか遅い。　pthreadだけ作業用にglobalでとってあげたりしたんだけどあんまり早くならない。　pthread_joinが結構食ってる感じだ。外すと他と同じぐらいになるのだけれど答えが合わない。　内積計算はメモリーアクセスの問題の方が大きいかも

研究室のクラスタでやったらXeon( sandy世代 )
 
     serial 0.19 s
     Omp   0.001 s ( 2 thread )
     pthread 0.001 s ( 2 thread )

もっと正確に測んなきゃだめですね。　とにかく順当に早くなっていました。Pthreadのほうが欠かなきゃいけない量が多いのでOpenMPですね。

EN
I just wondered which is faster OpenMP or Pthread. Here is my measurement.
The test code is illustrated below. It's a simple double precision dot product using cblas written in C. The compiler used was gcc-4.7.0. I initially tested on my Netbook with AMD C50 APU and got follwing result

real:
     serial 1.079 s
     Omp   0.830 s ( 2 thread )
     pthread 1.225 s ( 2 thread )

I didn't got what I expected somehow pthread was slower than serial ??? For validation, I also tested on my Uni's Cluster, which has sandy bridge generation Xeon, and got follwing result.

     serial 0.19 s
     Omp   0.001 s ( 2 thread )
     pthread 0.001 s ( 2 thread )

Looks OK now though more accurate measurement is required. Pthread looks slightly slower for me, yet pthread required more longer codes. Then I'm for OpenMP.

以下テストコード　
( serial, pthread and OpenMP performance comparison by Cblas dot product)

#include<stdio.h>
#include<stdlib.h>
#include<string.h>
#include<cblas.h>
#include<pthread.h>

#include<omp.h>

#define N 165535
//#define N 8

double *x_global;
double *y_global;
double tmp_global;

//--------------------------------------------------------
void init(
//--------------------------------------------------------
    double *x,
    double *y
){
    size_t i;
    for( i=0; i<N; i++ ){
        x[i] = i;
        y[i] = i;
    }
}

//--------------------------------------------------------
void serial_dot(
//--------------------------------------------------------
    double *x,
    double *y
){
    cblas_ddot( N, x, 1, y, 1 );   
    #ifdef SERIAL
    printf( " %lf\n", cblas_ddot( N, x, 1, y, 1 ) );   
    #endif
}

//--------------------------------------------------------
void omp_dot(
//--------------------------------------------------------
    double *x,
    double *y
){
    size_t i;
    const size_t nprocs = omp_get_num_procs();
    double tmp;

    omp_set_num_threads(nprocs);

    tmp = 0.0;   
    #pragma omp parallel for
    for( i=0; i<nprocs; i++ ){
        tmp += cblas_ddot( N/2, &x[i*N/2], 1, &y[i*N/2], 1 );
    }
    #ifdef OMP
    printf("%lf\n",tmp );
    #endif
}

//--------------------------------------------------------
void *pddot( void *arg ){
//--------------------------------------------------------
    size_t i;

    i = (size_t)arg;
    #ifdef PTHREAD
    printf("\tthread[%lu]\n",i);
    #endif

    tmp_global += cblas_ddot( N/2, &x_global[i*N/2], 1, &y_global[i*N/2], 1 );
}

//--------------------------------------------------------
void pthread_dot(
//--------------------------------------------------------
    pthread_t threads[2],
    double *x,
    double *y
){
    size_t i;

    tmp_global = 0;
    for( i=0; i<2; i++ ){
        pthread_create( &threads[i], NULL, pddot, (void*)i );
    }
    for( i=0; i<2; i++ ){
        pthread_join( threads[i], NULL );
    }
    #ifdef PTHREAD
    printf("%lf\n",tmp_global);
    #endif
}

//--------------------------------------------------------
int main(){
//--------------------------------------------------------
   
    int i;   
    pthread_t threads[2];
    double *x, *y;
    x = (double*)malloc(sizeof(double)*N);   
    y = (double*)malloc(sizeof(double)*N);   

    x_global = (double*)malloc(sizeof(double)*N);
    y_global = (double*)malloc(sizeof(double)*N);

    init( x, y );

    memcpy( x_global, x, sizeof(double)*N );
    memcpy( y_global, y, sizeof(double)*N );
   
    for( i=0; i<1000; i++ ){
        //serial_dot( x, y );
        //omp_dot( x, y );
        pthread_dot( threads, x, y );
    }

return 0;}