Falcon: Question: What WSGI server do you recommend with PyPy and does it outperform CPython in your testing?

[fgimian]

Hey there guys, firstly thanks for your awesome framework! 😄

I've been benchmarking Falcon on my end by running a WSGI server and using wrk to run some tests against a 'Hello world' endpoint.

What I'm finding is that CPython with gunicorn and Meinheld workers is by far the best option. I have tested out PyPy with uwsgi and gunicorn (with various types of workers) and can't get anywhere near the performance.

I have run your suite of benchmarks and can indeed replicate the incredible performance under PyPy but I wonder if the benchmarks are a bit unrealistic in the real world due to the omission of a WSGI server serving the traffic.

Are there any suggestions you have about getting the absolute best performance via PyPy and Falcon when serving via a WSGI server? Which server do you suggest and are there any recommended settings to squeeze every last bit of juice out of PyPy and Falcon?

Thank you so much!
Fotis

[vytas7]

Hello Fotis!
And thanks for the tips on https://github.com/falconry/falcon/issues/462 !

There is an in-progress pull request https://github.com/falconry/falcon/pull/1280 by @nZac for https://github.com/falconry/falcon/issues/462, however we have reduced its scope to NGINX+uWSGI for now to quicker get something usable out.

PyPy+Gunicorn documentation has been split out to a new issue: https://github.com/falconry/falcon/issues/1369 . Needless to say, contributions are very welcome on that one :slightly_smiling_face:

[vytas7]

Hi again @fgimian ,
I am going to close this issue by merging it with the mentioned #1369 .

I have run your suite of benchmarks and can indeed replicate the incredible performance under PyPy but I wonder if the benchmarks are a bit unrealistic in the real world due to the omission of a WSGI server serving the traffic.

A word on benchmarks. It is true that they are not representative of a "real" network server throughput -- the focus is instead on illustrating the overhead that different frameworks are incurring, sort of a theoretical maximum of what is possible, and when network is taken into account, the real throughput will be lower.

On the other hand, these benchmarks reflect numbers on a single vCPU core crunching Python code. Adding more worker processes on a multicore machine may actually get even higher numbers than these single-core benchmarks, as, for instance, seen on https://www.techempower.com/benchmarks/#section=data-r19&hw=ph&test=plaintext , on a physical multicore Xeon server (IIRC also using Meinheld as you suggest), one can get close to half million requests per second with Falcon.

[fgimian]

No worries at all @vytas7, thank you so much for the information too 🥘