Syscall overhead demonstration

20120512.c

#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <unistd.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <netinet/in.h>
#include <arpa/inet.h>
#include <time.h>
#include <string.h>

#define BUFFER_SIZE 128
#define NUM_ITERATIONS 5000000

void __attribute__((noinline)) append_and_send(int fd, unsigned char *buf1, unsigned char *buf2)
{
    memcpy(buf1 + BUFFER_SIZE, buf2, BUFFER_SIZE);
    send(fd, buf1, 2 * BUFFER_SIZE, 0);
}

static void send_twice(int fd, unsigned char *buf1, unsigned char *buf2)
{
    send(fd, buf1, BUFFER_SIZE, 0);
    send(fd, buf2, BUFFER_SIZE, 0);
}

static double timediff(const struct timespec *before, const struct timespec *after)
{
    return after->tv_sec - before->tv_sec + (after->tv_nsec - before->tv_nsec) / 1000000000.;
}

int main(int argc, char **argv)
{
    assert(argc == 3);

    struct sockaddr_in addr = { .sin_family = AF_INET, .sin_port = htons(atoi(argv[2])) };
    inet_aton(argv[1], &addr.sin_addr);

    int sockfd = socket(PF_INET, SOCK_STREAM, 0);
    assert(sockfd != -1);

    assert(connect(sockfd, (struct sockaddr *)&addr, sizeof(addr)) == 0);

    unsigned char buf1[BUFFER_SIZE * 2], buf2[BUFFER_SIZE];
    memset(buf1, 0xaa, sizeof(buf1));
    memset(buf2, 0x55, sizeof(buf2));

    struct timespec before, after;
    double t1, t2;

    clock_gettime(CLOCK_MONOTONIC, &before);
    for(size_t cnt = 0; cnt < NUM_ITERATIONS; ++cnt)
        append_and_send(sockfd, buf1, buf2);
    clock_gettime(CLOCK_MONOTONIC, &after);
    printf(" with memcpy: %.3f seconds\n", t1 = timediff(&before, &after));

    clock_gettime(CLOCK_MONOTONIC, &before);
    for(size_t cnt = 0; cnt < NUM_ITERATIONS; ++cnt)
        send_twice(sockfd, buf1, buf2);
    clock_gettime(CLOCK_MONOTONIC, &after);
    printf("with 2x send: %.3f seconds\n", t2 = timediff(&before, &after));
    
    printf("  difference: %.3fx\n", t2 / t1);

    close(sockfd);
}

output.txt

$ gcc -Wall -pedantic -std=c99 -O2 20130512.c -lrt

$ nc -l localhost 8888 >/dev/null & sleep 0.5; ./a.out localhost 8888
[1] 17687
 with memcpy: 2.020 seconds
with 2x send: 3.532 seconds
  difference: 1.748x
[1]+  Done                    nc -l localhost 8888 > /dev/null