These are chat archives for dropbox/pyston

5th
May 2015
Michael Arntzenius
@rntz
May 05 2015 01:25
ok, this is super weird. after rebasing onto master, the exceptions benchmark I made is like 6x slower
hm, nevermind, it's probably a cpu frequency scaling issue
Michael Arntzenius
@rntz
May 05 2015 01:30
ha, it was totally a cpu frequency scaling issue
Chris Toshok
@toshok
May 05 2015 01:30
ugggh
don’t say those words :)
Michael Arntzenius
@rntz
May 05 2015 01:30
turns out going from 600 MHz to 2.5 GHz will really change your performance characteristics
Michael Arntzenius
@rntz
May 05 2015 03:37
nice! configuring libunwind with --disable-conservative-checks --disable-block-signals shaves 20% off the exception microbenchmark
now, I have no idea whether that's actually safe or not
Travis Hance
@tjhance
May 05 2015 03:37
‘block signals’?
Michael Arntzenius
@rntz
May 05 2015 03:39
libunwind calls sigprocmask() in various places to block signals. I'm not sure exactly why - which is why I'm not sure that disabling it is safe. But it removes a bunch of syscalls from the unwinder.
I think they're concerned about a signal causing libunwind to be entered recursively
Marius Wachtler
@undingen
May 05 2015 08:35
looks like we conservatively scan the memory of the weakref.proxy object and find there the pointer to the referenced object. turning the weak ref in a not so weak one... :-D
Marius Wachtler
@undingen
May 05 2015 22:15
has anyone took a look at: http://pypyjs.org/ perf: http://arewepythonyet.com/ ? :-D

PyPy.js is an experiment in building a fast and compliant python environment for the web.

It uses the PyPy python interpreter, compiled for the web via emscripten, with a custom JIT backend that emits asm.js code at runtime.

Kevin Modzelewski
@kmod
May 05 2015 22:18
heh :)
Chris Toshok
@toshok
May 05 2015 22:39
wow, INSTALLING.md is so much nicer these days
Marius Wachtler
@undingen
May 05 2015 22:42
before one had to be really curious about pyston to follow thru all the steps...
Kevin Modzelewski
@kmod
May 05 2015 22:46
haha yeah much better now
Chris Toshok
@toshok
May 05 2015 23:17
ah, the BoxedDict size test was failing because it’s a c file, being compiled with clang, not clang++