Slow accept() in socket server in C - c

Question

I have the following code for my education socket server in C.
#include <arpa/inet.h>
#include <netinet/in.h>
#include <stdio.h>
#include <sys/socket.h>
#include <sys/types.h>
#include <time.h>
#include <unistd.h>
double get_wall_time()
{
struct timeval time;
if (gettimeofday(&time, NULL)){
return 0;
}
return (double)time.tv_sec + (double)time.tv_usec * 0.000001;
}
double get_cpu_time()
{
return (double)clock() / CLOCKS_PER_SEC;
}
int main()
{
double wall = get_wall_time();
double cpu = get_cpu_time();
int sfd = socket(AF_INET, SOCK_STREAM, 0);
struct sockaddr_in own_addr = {0};
own_addr.sin_family = AF_INET;
own_addr.sin_port = htons(5678);
bind(sfd, (struct sockaddr *)&own_addr, sizeof(own_addr));
listen(sfd, 5);
static char message[] = "hello from server\n";
double wall_accept = 0;
double cpu_accept = 0;
int count = 0;
while (1) {
if (count++ == 1000) {
break;
}
double wall_start = get_wall_time();
double cpu_start = get_cpu_time();
int client_sfd = accept(sfd, NULL, NULL);
wall_accept += get_wall_time() - wall_start;
cpu_accept += get_cpu_time() - cpu_start;
send(client_sfd, message, sizeof(message), 0);
close(client_sfd);
}
wall = get_wall_time() - wall;
cpu = get_cpu_time() - cpu;
printf("wall accept: %lf\n", wall_accept);
printf("cpu accept: %lf\n", cpu_accept);
printf("wall: %lf\n", wall);
printf("cpu: %lf\n", cpu);
}
To test I use seq 1000 | time parallel -j 1 -n0 'nc 127.0.0.1 5678' | wc -l with results
wall accept: 6.436480
cpu accept: 0.010000
wall: 6.456266
cpu: 0.020000
For 10000 requests result is
wall accept: 55.434541
cpu accept: 0.080000
wall: 55.633679
cpu: 0.260000
Is accept() slow or I do something wrong? Or maybe this is normal result for single-thread implementation?
UPD. I also wrote a server with pthreads to send a message in different thread.
#include <arpa/inet.h>
#include <netinet/in.h>
#include <stdio.h>
#include <sys/socket.h>
#include <sys/types.h>
#include <time.h>
#include <unistd.h>
#include <pthread.h>
#include <stdlib.h>
#include <sys/time.h>
double get_wall_time()
{
struct timeval time;
if (gettimeofday(&time, NULL)){
return 0;
}
return (double)time.tv_sec + (double)time.tv_usec * 0.000001;
}
double get_cpu_time()
{
return (double)clock() / CLOCKS_PER_SEC;
}
void *send_message(void *pclient_sfd)
{
int client_sfd = *(int *)pclient_sfd;
free(pclient_sfd);
static char message[] = "hello from server\n";
send(client_sfd, message, sizeof(message), 0);
close(client_sfd);
return NULL;
}
int main()
{
double wall = get_wall_time();
double cpu = get_cpu_time();
int sfd = socket(AF_INET, SOCK_STREAM, 0);
struct sockaddr_in own_addr = {0};
own_addr.sin_family = AF_INET;
own_addr.sin_port = htons(5678);
bind(sfd, (struct sockaddr *)&own_addr, sizeof(own_addr));
listen(sfd, 5);
double wall_accept = 0;
double cpu_accept = 0;
int count = 0;
while (1) {
if (count++ == 10000) {
break;
}
int *pclient_sfd = malloc(sizeof(*pclient_sfd));
double wall_start = get_wall_time();
double cpu_start = get_cpu_time();
*pclient_sfd = accept(sfd, NULL, NULL);
wall_accept += get_wall_time() - wall_start;
cpu_accept += get_cpu_time() - cpu_start;
pthread_t tid;
pthread_create(&tid, NULL, send_message, (void *)pclient_sfd);
}
wall = get_wall_time() - wall;
cpu = get_cpu_time() - cpu;
printf("wall accept: %lf\n", wall_accept);
printf("cpu accept: %lf\n", cpu_accept);
printf("wall: %lf\n", wall);
printf("cpu: %lf\n", cpu);
return 0;
}
Then I use seq 10000 | time parallel -j 4 -n0 'nc 127.0.0.1 5678' | wc -l and it takes 58 seconds.

Answer 1

It's the way you're testing it. When you use this
seq 10000 | time parallel -j 4 -n0 'nc 127.0.0.1 5678' | wc -l
That's actually going to impact the test because you're spawning lots of processes etc ie you're not actually testing the C application your testing the ability to spawn processes.
If we change this to a simple python script ie
#!/usr/bin/env python
import socket
TCP_IP = '127.0.0.1'
TCP_PORT = 5678
BUFFER_SIZE = 1024
msg = "1"
for i in range(0, 1000):
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
s.connect((TCP_IP, TCP_PORT))
s.send(msg.encode('utf-8'))
data = s.recv(BUFFER_SIZE)
s.close()
print("received data: %s" % data)
and run the test the results are vastly different.
real 0m0.269s
user 0m0.074s
sys 0m0.114s
If you really want to test this and see how fast it is you need to use separate machines etc and typically you might need to use C or your limiting factor might be the client. The best tool I've seen for this (HTTP Specific) where you might find some good code is wrk
https://github.com/wg/wrk

Answer 2

Take a look at your listen() second argument. Try to increase it.
here is text from man 2 listen
The backlog argument defines the maximum length to which the queue of
pending connections for sockfd may grow. If a connection request
arrives when the queue is full, the client may receive an error with
an indication of ECONNREFUSED or, if the underlying protocol supports
retransmission, the request may be ignored so that a later reattempt
at connection succeeds.