Does anyone know why JIT-ing Python isn't given more "institutional" support (by which I suppose I mean mostly Google, since they employ Van Rossum after all)? It seems like the logical next step to make Python more generally useful.

As I see it, speed is the last big advantage Java has over Python (and I do realize that a dynamically typed language makes achieving efficiency harder).

Apparently PyPy has some financial support from Google (http://en.wikipedia.org/wiki/Pypy#Project_status), but it really seems like basically any company that uses Python would gain so much from robust JIT compilation in the interpreter that I'm surprised there's not a more concerted effort to make it happen.

You bring up a good point. It bothers me a great deal that companies aren't actively supporting the software that they depend on. The Python Foundation will take your donations. A thousand dollars a year is peanuts and the dividends are huge.

Maybe because raw excution speed matters so little in most applications that they're rather focussing on lower hanging fruits such as improvements to the core language or standard library. - But that's just a wild guess.

Raw execution speed does matter, but like any incremental cost, is often ignored.

Off-the-cuff benchmarks I ran on our Scala-based webapp on a four-core desktop system demonstrated 5,000 req/sec @ 2ms/request after the JIT warmed up.

5,000+ requests/second -- scaling up with available CPUs -- means not having to worry about performance, complicated caching, and the slew of other things that developers do to eek performance out of frameworks based on poor interpreter implementations.

It's not really an incremental cost but rather a constant overhead.

The idea is that if you're worried about performance and scaling then you'll have to find a way to distribute your load over multiple physical machines anyways.

At that point it doesn't matter so much anymore whether one of your nodes handles 5000 reqs/sec or 2500 reqs/sec. Hardware is cheap.

"Hardware is cheap" is a false dichotomy. The implication is that hardware is cheap, but development is not, and performant architectures require expensive development.

In reality, neither hardware or development are cheap, and performance does not inherently necessitate more costly development. The mantra and the dichotomy are false.

Raw execution speed matters when dealing with Google sized data sets, but Google and organizations with similar requirements can just throw more hardware at it.

Google uses C++ for compute-bound stuff.

Google's problems are big enough that constants matter.

Cool to hear that things are moving on. I'm expecting great things from PyPy :)

Psyco is pretty cool right now, even without generator support (though that would be good).

I'm doing stuff with a neural net built in Python/Numpy right now and Psyco makes it twice as fast. Given that I'm waiting for it to run each time I try something this sort of speed up is quite nice. It's now to the point where pretty much all the execution time is spent in Numpy (which is largely implemented in C) which is cool.

The main problem with psyco is lack of 64-bit support. While a new psyco would be nice, what really would make sense is a boost-like tophat on rpython so you can drop into it and compile your performance-critical routines. Today people do that with C++ and boost. Cython/pyrex don't cut it because they don't allow you to create C-style strongly typed objects with Python-accessible methods.

Maybe you can do that today, but good luck finding out by reading the pypy docs.

I recently wrote some image-manipulation routines in Python which ran 71 times faster with Psyco. Without Psyco, this kind of computation isn't really feasible in pure Python, so I'm also a little surprised Psyco hasn't had more influence on C Python. But perhaps its just a matter of time before PyPy is bequeathed Psyco optimizations.