How do I reduce the time of context switching after waking from sem_wait

I found the time gap of sem_post and sem_wait is at least 50 microseconds.

sem_t gSema;
struct timeval gTv;

void *run(void *arg) {
    int result;
    struct timeval tv;
    struct timespec nano = {0, 1};

    while( 1 ) {
        result = sem_trywait(&gSema);
        if(result < 0) {
            nanosleep(&nano, NULL);
            continue;
        }
        if(result == 0) {
            gettimeofday(&tv, NULL);
            printf("waken up...elapsed time = %ld\n", (tv.tv_sec - gTv.tv_sec) * 1000000 + (tv.tv_usec - gTv.tv_usec));
        }
    }

    return NULL;
}
void *run2(void *arg) {
    int result;
    struct timeval tv;

    while( 1 ) {
        result = sem_wait(&gSema);
        if( result == 0 ) {
            gettimeofday(&tv, NULL);
            printf("waken up...elapsed time = %ld\n", (tv.tv_sec - gTv.tv_sec) * 1000000 + (tv.tv_usec - gTv.tv_usec));
        }
    }
    return NULL;
}
int main(int argc, char **argv) {
    pthread_t thr;
    sem_init(&gSema, 0, 0);
    pthread_create(&thr, NULL, run, NULL);

    while( 1 ) {
        sleep(1);
        gettimeofday(&gTv, NULL);
        sem_post(&gSema);
    }
    return 1;
}

At first, when I run the code with run2(), elapsed time is 50 microseconds. Then CPU usage is very low. Second, when I run the code with run() without nanosleep, elapsed time is 0 or 1. But CPU usage is 100% expectably. With nanosleep, elapsed time is 1~50 microseconds.

Is there some better method of low cpu usage AND low latency?