Well here goes. My site is this http://codeflow.org/ I haven't been terribly happy with Apache, it's, fast-ish, but configuring it is pain. Also dispatching dynamic requests via (whatever mod-wsgi or something) isn't quite as fast. I've run cherokee for the last couple of years, but have my own grievances with that (doesn't support gzip+content-length, impossible to hand edit configuration, kinda slow, that kind of thing again).

I've had run-ins with nginx, and it's pretty cool, except configuring it is a giant pain. Also, it still isn't dispatching dynamic requests as fast as could be (goes trough a socket to another process or whatever).

Then there's stuff that no webserver out there does really satisfactory, which is low-latency websockets. Reason is that they proxy the websockets trough to another process, and even though it's on the local machine, this just drives latency up by whole dozens of milliseconds...

Ok, so what's my obsession with speed all about? My capital is strictly limited, so I've got to run whatever I want to run on my own VPS. Unfortunately I started running http://webglstats.com/ the tracker of which is embedded in some 500 other sites. They shuffle traffic me at a rate of between 5-10 stat impressions per second, and each of those triggers 3 requests (two gets, one post). I concurrently run around 50-150 open connections at all times.

So naturally I'd want the tracker to run as fast as I can, and I'd like my other web properties not to slow down because I run a high-traffic service on that server. And I found that pretty much everything out there just isn't very good at that.

So having had finally enough after years, I wrote my own. It's about 1500 lines of python, it's asynchronous, can do http-pipelining, has a dynamic dispatch router that can do things like vhosts, has an inotify integration so it avoids stat'ing static files all the time, it lets me plug python modules directly into the webserver (so no more subprocesses/sockets/fastcgi/whatever for dynamic requests). It does around 4000 requests/s, it easily does at least 512 concurrent open connections, it peaks at a troughput on localhost of +300mb/s and it does that while maintaining a maximum latency (on localhost) of 2ms, or outside (across about 6 hops) of around 15ms.

I'm happy with that, and now that the I spent the last two weeks writing that server (it's not the first I wrote, but I finally took the leap), I'm 100% sure that whatever problem I encounter in my hosting adventures, I can make my server solve them.