[Libevent-users] performance patch

[Vsevolod Stakhov]

Tani Hosokawa пишет:
> Under high load, unless you build a dedicated acceptor thread you will 
> find that only allowing a listen backlog of 10 connections will cause 
> lots of attempted connections to simply be dropped if you have any 
> significant amount of processing to do per request.  The reason I put 
> 1024 in there is all implementations that I know of silently truncate 
> the backlog parameter to the system's maximum.  The reason I don't use 
> SOMAXCONN is some OS (older Solaris for sure) are incorrect for backward 
> compatibility reasons, and other OS are incorrect because the actual 
> setting is dynamically tunable via sysctl.
> 
> --- http.c.old  2008-02-26 12:30:45.000000000 -0800
> +++ http.c      2008-02-26 12:33:54.000000000 -0800
> @@ -2002,7 +2002,7 @@
>        if ((fd = bind_socket(address, port)) == -1)
>                return (-1);
> 
> -       if (listen(fd, 10) == -1) {
> +       if (listen(fd, 1024) == -1) {
>                event_warn("%s: listen", __func__);
>                EVUTIL_CLOSESOCKET(fd);
>                return (-1);
> 

As POSIX requires SOMAXCONN to be defined in sys/socket.h it is safe to 
use this value or define some constant in configure script, that detects 
  SOMAXCONN definition. Also in old Solaris systems backlog was limited 
to 5.