These are chat archives for dropbox/pyston

13th
Feb 2016
Marius Wachtler
@undingen
Feb 13 2016 14:10
interesting, haven't heard of that allocator yet. But I don't think that switching away from jemalloc would change performance much at this point. (Because our GC is using it's on memory management which is faster than jemalloc) After the refcounting switch we may see larger perf changes with different allocators. :-)
Chris Seaton
@chrisseaton
Feb 13 2016 18:53
@undingen only one thing to do: make allocators swappable and run an experiment!
Kevin Modzelewski
@kmod
Feb 13 2016 21:18
It's interesting that we can beat jemalloc
I also noticed that in the refcounting branch (where everything gets allocated with malloc), jemalloc didn't seem particularly amazing
I wonder what they're having to do that we avoid with a custom allocator
Chris Seaton
@chrisseaton
Feb 13 2016 21:19
Do you do any allocation removal yet?
Kevin Modzelewski
@kmod
Feb 13 2016 22:01
only some very simple cases
One issue is that the program can at any time reach for the variable indirectly, such as sys._getframe(0).f_locals['my_var_name']
so if we remove the allocation, we probably have to be prepared to rematerialize it when something like that happens
Chris Seaton
@chrisseaton
Feb 13 2016 22:06
Yes we have the same problem in Ruby, so we rematerialise on demand