500,000 Requests/Sec – Modern HTTP Servers Are Fast

Sigh, another completely synthetic benchmark with big iron: A dual Intel Xeon X5670 with 24GB of RAM from SoftLayer. The X5670 has 6 cores @ 2.93 GHz, 2 threads per core, /proc/cpuinfo shows 24 CPUs.<p>Serving a plain HTML page over <i>localhost</i> via nginx. A medium length blog post with pretty much no real-world information.

My toy web server can achieve similar performance using a much more modest hardware (Core i7 2640 laptop), using way less RAM (a few dozen kilobytes).<p>Granted this is also being tested on localhost with static content (no disk I/O) -- but shows that event-driven servers are not that novel or difficult to write: my code weighs around 1700 LOC of (might I say) readable C.<p>Static file serving is also fast (using sendfile(), etc), but needs an overhaul to achieve usable concurrency. Currently there's a ~4x performance drop while serving files, but I'm working on this.<p>(The sources are at <a href="http://github.com/lpereira/lwan" rel="nofollow">http://github.com/lpereira/lwan</a> by the way.)

nginx saturates at ~18k req/sec / core with a latency of ~24ms. This saturation is not coming from nginx in particular but from OS limits (mode switching, stack copying etc.. for read/write system calls).<p>There is nothing new in "modern HTTP servers". They are event-driven programs and this has existed for a long time.

In a real world app, I think you'd be considered lucky if the web-server is your bottleneck, rather than your DB or network connection.

Quickly looking over SoftLayer's price sheet, it looks to be a little over $1k a month for Intel 5670 with 24GB of RAM

Since this is almost pure sendfile() work (aside from the headers), it really doesn't seem like a very useful example... there isn't much static content left on the web.

The flipside is that as time goes on, this benchmark becomes less impressive. The server in question had 24GB of RAM and two CPUs with 12 cores each.

A comparison with older hardware would have been nice. Also using localhost is not a very good indicator of real-world performance.

The ironic part is that his blog is hosted on wordpress.

I'm interested in more info on your linux TCP tuning, especially how you decided to tcp_tw_recycle and tcp_fin_timeout (and how you decided setting the former to 1 is safe)