NUMA memory access latency benchmark

/* My first pthread routine to test the local and remote memory access time difference of AMD and Intel CPU*/
 
#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <sys/time.h>
#include <unistd.h>
#include <sched.h> 
#include <pthread.h>
#include <errno.h>
#define handle_error_en(en, msg) \
               do { errno = en; perror(msg); exit(EXIT_FAILURE); } while (0)
#define MEM_SIZE 1024*1024*1024
#define NUM 10
cpu_set_t cpu_info;
pthread_barrier_t   barrier;
pthread_barrier_t   barrier1;
int *b;
void *thread()
{
        int s,i,j;
        pthread_t myid = pthread_self();
        CPU_ZERO(&cpu_info);
        CPU_SET(1, &cpu_info);
        s = pthread_setaffinity_np( myid  , sizeof(cpu_set_t), &cpu_info);
        /*if(ret_v != 0){printf("New thread set failed\n"); exit(1);}
        ret_v = sched_getaffinity( myid  , sizeof(cpu_set_t), &cpu_info_get);
        if(ret_v != 0){printf("Main thread get failed\n"); exit(1);}
        for(i=0; i<sizeof(cpu_set_t); i++) 
         printf("%d", cpu_info_get.__bits[i]);*/
        if (s != 0)
        handle_error_en(s, "pthread_setaffinity_np");
           /* Check the actual affinity mask assigned to the thread */
        s = pthread_getaffinity_np(myid, sizeof(cpu_set_t), &cpu_info);
        if (s != 0)
             handle_error_en(s, "pthread_getaffinity_np");
        printf("Set returned by pthread_getaffinity_np() contained:\n");
        for (j = 0; j < CPU_SETSIZE; j++)
               if (CPU_ISSET(j, &cpu_info))
                   printf("    CPU %d\n", j);
        printf("\nHello! New thread!\n");
        //pthread_barrier_wait(&barrier);
        b = (int*)malloc(sizeof(int) * MEM_SIZE);
        if(b == NULL) {printf("Malloc failed!\n"); exit(1);}
        for(i=0; i<MEM_SIZE; i++)
                b[i]= 0;
        printf("malloc and touch b finished!\n");
        pthread_barrier_wait(&barrier);
        pthread_barrier_wait(&barrier1);
        free(b);
        pthread_exit( (void*)1 );
}
int main()
{       
        pthread_t pt;
        int s,ret_v,i,j;
        pthread_t myid = pthread_self();
        pthread_barrier_init (&barrier, NULL, 2);
        pthread_barrier_init (&barrier1, NULL, 2);
        printf("Staring main....\n");        
        CPU_ZERO(&cpu_info);
        CPU_SET(0, &cpu_info);
        /*
        ret_v = sched_setaffinity(0, sizeof(cpu_set_t), &cpu_info); 
        if(ret_v != 0){printf("Main thread set failed\n"); exit(1);}
        ret_v = sched_getaffinity(0 , sizeof(cpu_set_t), &cpu_info);
        if(ret_v != 0){printf("Main thread get failed\n"); exit(1);}
        for(i=0; i<sizeof(cpu_set_t); i++)
                printf("%d", cpu_info.__bits[i]);
        putchar('\n');*/
        s = pthread_setaffinity_np( myid  , sizeof(cpu_set_t), &cpu_info);
        if (s != 0)
            handle_error_en(s, "pthread_setaffinity_np");
           /* Check the actual affinity mask assigned to the thread */
        s = pthread_getaffinity_np(myid, sizeof(cpu_set_t), &cpu_info);
        if (s != 0)
             handle_error_en(s, "pthread_getaffinity_np");
        printf("Set returned by pthread_getaffinity_np() contained:\n");
        for (j = 0; j < CPU_SETSIZE; j++)
               if (CPU_ISSET(j, &cpu_info))
                   printf("Main thread on    CPU %d\n", j);
       
        int *a = (int*)malloc(sizeof(int)*MEM_SIZE);
        if(a == NULL) printf("Malloc failed!\n");
        double timeuse=0.0;
        for(i=0; i<MEM_SIZE; i++)
                a[i]=0;
        printf("Memory a touched!\n");
        ret_v = pthread_create(&pt,NULL,thread,NULL);
        if(ret_v !=0 ){printf("Thread creat error!"); exit(1);}
        printf("New thread created!\n");
        /*CPU_ZERO(&cpu_info);
        CPU_SET(1, &cpu_info);
        ret_v = sched_setaffinity( pt, sizeof(cpu_set_t), &cpu_info);
        printf("New thread set!\n");*/
        pthread_barrier_wait(&barrier);
       
        printf("######################\n");
        struct timeval start;
        struct timeval end;
        srand(6);
        gettimeofday(&start, NULL);
        for(j=0; j<NUM; j++)
         for(i=0; i<MEM_SIZE; i++)
                 a[i]=j;
        gettimeofday(&end, NULL);
        timeuse = 1e6*(end.tv_sec - start.tv_sec) + end.tv_usec - start.tv_usec;
        printf("Local Memory time %lf  us \n",(timeuse/NUM));
        printf("######################\n");
        srand(6);
        gettimeofday(&start, NULL);
        for(j=0; j<NUM; j++)
        for(i=0; i<MEM_SIZE; i++)
                 b[i]=j;
        gettimeofday(&end, NULL);
        timeuse = 1e6*(end.tv_sec - start.tv_sec) + end.tv_usec - start.tv_usec;
        printf("Remote Memory time %lf  us \n",(timeuse/NUM));
        pthread_barrier_wait(&barrier1);
        pthread_barrier_destroy(&barrier);
        pthread_barrier_destroy(&barrier1);
        free(a);
        return 0;
}