News for hackers. Stuff that matters.
Re:Benchmark of Python web servers by ubernostrum on Tuesday March 16, 2010 @11:56AM (#1195964)
Unfortunately there are several problems being pointed out with this benchmark:

* Some of the servers tested aren't designed to interact directly with clients over a network, and are meant instead to run proxied over a local socket behind something else like nginx. The benchmark is inconsistent in whether it actually does this, though, which makes the numbers useless for comparison purposes.

* Several of these servers use preforking architectures which assume the ability to have multiple workers available to service requests. Yet the benchmark deliberately limited them to only one worker, which makes the numbers useless for gauging actual peformance under concurrent load.

Re:Benchmark of Python web servers by ubernostrum on Tuesday March 16, 2010 @06:47PM (#1197028)
And now there are updated numbers in the graphs (labeled "gunicorn-3w") for gunicorn, based on running it the way it's meant to be run (on a socket with more than one worker). The results:

* Concurrent requests more than doubled

* Response times dropped 75%

* Error rate dropped 75%

It's amazing what a difference it makes to actually run something the way it was designed to run...

Re:Benchmark of Python web servers by davisp on Tuesday March 16, 2010 @08:33PM (#1197278)
I'm glad he updated it. A lot of people would've just left it. That said I also would've liked to see at least a note that he tried more than three workers and there wasn't an improvement. When you have an almost linear improvement its a bit confusing why you wouldn't keep increasing it.