Pyston 2.1 Is Blowing Past Python 3.8/3.9 Performance

[zxy_thf]

50% is impressive, but not good enough for production.
Like what Dropbox has done, replacing performance critical part with compiled language (C/C++/Go/Rust) is more practical and efficient.

I think for the most ML programs, python is just an interface while the core is written in C/C++.

It's not only ML. This is the standard practice for any performance-demanding python package. (numpy, scipy, etc)

[rmfx]

because a lot of python extensions don't work with pypy. it would break a lot of code.

I didn’t say it could replace it tomorrow, i said they should starts preparing a cpython replacement with pypy (and complete it with what’s needed for this).

[Raka555]

I wonder how tcc as a JIT compiler would perform vs python, ruby, javascript, etc.

The code that tcc produces is not very optimised, that is why the compile time is so fast.

[lyamc]

The code that tcc produces is not very optimised, that is why the compile time is so fast.

Yeah, I know. I want to know how does it compare to python in execution time if it were used as a JIT compiler.

[tildearrow]

I wonder how [URL="https://github.com/run4flat/C-TinyCompiler"]tcc as a JIT[URL]

Dang it, I am waiting until the day Python gets a JIT... just like most Java VMs...

[ssokolow]

Dang it, I am waiting until the day Python gets a JIT... just like most Java VMs...

That's what PyPy is. Improvements to it have levelled off around 4.25x CPython speed because Python, as a language, has a level of dynamism to it that's inherently difficult to optimize.

(And because some of the benchmarks that get geometric-averaged together really don't make up the difference after taking the hit from the overhead of the runtime figuring out what to JIT. For example, sphinx and sympy_integrate slam into the "150% of CPython runtime" mark at the top of the graph.)

[caligula]

50% is impressive, but not good enough for production.
Like what Dropbox has done, replacing performance critical part with compiled language (C/C++/Go/Rust) is more practical and efficient.

What Dropbox is doing isn't that impressive. At least the Linux client is huge. It's almost as large as the whole Nextcloud suite (server side), which contains tons of functionality.

[coder]

No source, no interest.

Life is too short to be bogged down by code you can't fix yourself.

While I agree that I wouldn't use a closed-source solution, I still appreciate having benchmarks against it.

[coder]

That's what PyPy is. Improvements to it have levelled off around 4.25x CPython speed because Python, as a language, has a level of dynamism to it that's inherently difficult to optimize.

Excuse my ignorance, but does it perform any form of profile-driven optimization? Wikipedia mentions that it uses a "tracing JIT", which seems to imply some level of feedback.

[Paradigm Shifter]

Nice to see some benchmarks. It's a fairly impressive jump. I would be happier when they release the source code, though.