[enoptix]

Much better memory usage and (for us at least) better concurrency. We could only run 4 unicorn workers on a single dyno. But with Puma we run 16 threads with ease.

[nixme]

Have you tried running 3 or even 4 puma workers (each with 8-16) threads on a dyno? That way you can get more concurrency on CPU-bound requests in addition to IO-bound concurrency (assuming MRI).

[jrochkind1]

I tried that in my attempt at benchmarking a standardized app on heroku, and, yep, it works great:

https://github.com/jrochkind/fake_work_app/blob/master/READM...

[cpursley]

Thanks, this is all very relevant at the moment. Trying to optimize page load times for a Heroku Rails app. Web server is one of the bottlenecks.