Just a followup to my previous thread, but I think it can interest
some developpers.
Nginx is great! I’ve tried before to “hack” Lighttpd, but it was a
pain in the a*s. It takes me some times to enter Nginx but I don’t
regret it at all. Thanks Igor for this superb work.
So, I was trying to call (long) blocking functions inside Nginx.
First approach has been to create worker threads, a queue and a
thread which is doing the epoll_wait and put in the queue
epoll related events as well as my workers events. It was fine
but a bit complicated and performance was around 10000
requests / s (a dynamic 10 kB document).
Clearly:
static request → nginx → msg → nginx
dynamic request → nginx > msg → my worker → msg → nginx
I’ve just discovered signalfd(). It’s now (with a recent Linux
kernel) possible to associate a file descriptor with a signal
handler and it’s compatible with epoll_wait 
so now it’s:
static request → nginx
dynamic request → nginx → msg → my worker → signal → nginx
So in the event loop, it’s like :
nevents = epoll_wait (…event_list…) ;
for (i = 0 ; i < nevents ; i ++)
{
if (event_list [i].data.fd == my_fd)
{
process_my_eventS (…) ;
continue ;
}
…
}
process_my_eventS (…)
{
msg_t msg ;
while (1)
{
poll_msg (q, & msg) ;
if (read (myfd…) != sizeof (signalfd_siginfo))
break ;
handle_http_results (msg.request) ;
}
}
Of course I still need a queue to get the context of the request,
but it works fine (100.10^6 requests so far without a glitch
using ab).
From my benchmarks it’s now ~ 15500 requests / s (10 kB dynamic
document).
Just to compare:
Apache 2.2 / Prefork MPM / PHP 5 with index.php:
<?php echo '<10kB of ascii data>' ; ?>1800 requests / s, 80 load!
And with Nginx:
15500 requests / s, ~ 2 load 8)
Using ab -k -c 50 -n 1000000 http://127.0.0.1:8000/
On an AMD X2 4200 / 2 GB / Ubuntu 8.04 RC1.
Sorry to bother you with this long message…