This is definitely interesting, but what I'm missing is comparison with a single dyno running Thin.

Unicorn (3 workers) + Sinatra gets 2,300 requests per second.

Thin + Sinatra gets 1,000 requests per second.

https://github.com/joevandyk/unicorn-on-heroku for unicorn code/results.

https://github.com/joevandyk/unicorn-on-heroku/tree/thin for thin code/results.

The weird thing is I tried scaling up to 10 dynos (even tried 50). I didn't see any throughput improvement.

Just curious, where did you execute ab from? I've uploaded your code and ran ab from my laptop - I've tried it under 3 different internet connections - and I only get about 100 req/s.

Thanks for testing and for sharing the results.

Thanks for testing it out! Great results!

If your app is doing any network IO, I'd expect unicorn + multiple worker processes to give even better results over thin.

1 Unicorn worker is roughly equivalent to a Thin instance. You can only put a single Thin instance on a dyno, meaning that the first benchmark is what you previously could only attain in terms of throughput.

I might've not explicitly said this in my article, but see the first benchmark where I say:

"1 Unicorn Worker (Roughly equivalent to a single Thin instance)" - stating that this is Thin's (and was Heroku's) previous limit.

With Unicorn you can gain 3 to 4 times more throughput.

Thanks for clarifying. The reason I asked, is that I thought maybe Heroku did some optimizations on their stack that uses Thin.

No problem. I doubt there is anything app-server specific to the Celadon Cedar stack. Maybe the Bambidious Bamboo stacks there is Thin-specific configuration. But Cedar is meant to run anything you want through the Procfile, you could also use Mongrel or Rainbows or maybe even Passenger. But Unicorn gets my preference as of late! Amazing app server.