[Scons-users] Scons on pypy

[Russel Winder]

On Sun, 2016-06-05 at 14:15 +0200, Dirk Bächle wrote:
> […]
> 
> If you'd like to raise attention to this topic again, I'd like to
> hear what the background of your usage scenario is. Has pypy added 
> some changes recently that seem like they could offer a substantial
> running time advantage over CPython? We'd certainly be 
> interested to hear about that...
> […]

Someone mentioned an execution speed up of 50x using PyPy compared to
CPython. I can attest that for computationally intensive (CPU bound)
things, PyPy is indeed 5x to 30x faster than CPython. For I/O bound
there is effectively no difference at all.

I suspect, but have no data of any sort, that the nature of the SCons
code is that the I/O bound dominates the CPU bound, thus I have never
pushed the PyPy angle for performance. Also, given the lack of serious
backing of PyPy by people who could easily afford the resource and rely
on fast Python execution, it is going to remain a very minor player.

So what about Python performance? Well there is:

– NumPy but I suspect the available data structures and algorithms are
of little use to SCons.

– Cython but is it easy to highlight the performance critical part of
SCons that can be converted to static code. And anyway Cython is not
the right technology these days…

– Numba  but is it easy to highlight the performance critical part of
SCons that can be compiled.

– D, Rust, Go, C++, Chapel, i.e. use a polyglot approach with the CPU
intensive part in native code from the outset.

To be honest data science is staying with NumPy despite its problems
(probably due to SciPy and especially Pandas) and maybe doing a bit of
Numba (because it is cool and amazingly good use of LLVM), but the real
people doing HPC are doing things with D and Chapel.

I wonder if rewriting SCons in D would eliminate all the performance
issues, I suspect yes.

All in all 7 or 8 years ago I was a huge fan of PyPy and did various
things behind the scense to try and get it funded. However now in 2016,
I think it is effectively an interesting backwater. My view re PyPy got
cemented after a long conversation with Larry Hastings at PyConUK 2015.

-- 
Russel.
=============================================================================
Dr Russel Winder      t: +44 20 7585 2200   voip: sip:russel.winder at ekiga.net
41 Buckmaster Road    m: +44 7770 465 077   xmpp: russel at winder.org.uk
London SW11 1EN, UK   w: www.russel.org.uk  skype: russel_winder
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part
URL: <https://pairlist4.pair.net/pipermail/scons-users/attachments/20160605/453f0e21/attachment.pgp>