The nginx default settings are a stroke of accedental or maybe intentional genius. They are so under powered that it is actually amazing they worked for me for as long as I had them. I assume it is this way so that it forces admins to actually pay attention and tune it for production environments. I run an open bittorrent tracker that tracks somewhere upwards of several million peers concurrently and I use nginx as my load balancer for the http side of it.
everything is fire
a few days ago i noticed that the box i was running this massive open tracker on was dogpiling and had a backlog of over 1K connections in the workers in the writing state. after putting nginx performance tuning into the web search, i came to the realization that you are actually supposed to tune nginx for production enviroments because the defaults are so .... low end? you could probably run nginx on a netbsd lemon so i guess that's fine.
by default nginx processes 1 connection per worker.
to ammend this, set mulit_accept in the events block of your nginx config to be on as god intended. while i was in the config, i also made other numbers go up:
events {
multi_accept on;
worker_aio_requests 1024;
worker_connections 10240;
}
additionally i increased the open file limits in nginx's system unit to a comically huge number, 10 million (dont do this, use a smaller number).
# note: this is not 10 million. this is a sane value that people will not copy paste into their configs.
[Service]
LimitNOFILE=10000
number go up haha
as a result, munin graphs showed that indeed number go up as a result of using non default nginx configs.
you can see were the tuning occured pretty clearly.
