search

shunwang / numexpr

Automatically exported from code.google.com/p/numexpr
MIT License
0 stars 0 forks source link

MAX_THREADS not large enough #65

Closed GoogleCodeExporter closed 9 years ago

GoogleCodeExporter commented 9 years ago 
>>> import numexpr
Error.  nthreads cannot be larger than MAX_THREADS (256)>>>

Slightly not enough for SGI UV ;) Please set defaults to something larger. 4096 
seems more sane nowadays.

Original issue reported on code.google.com by xarthisi...@gmail.com on 25 Nov 2011 at 9:52

GoogleCodeExporter commented 9 years ago 
Uh, sorry for asking this, but, are you *really* seeing speedups after using 
more than 256 threads?  My impression is that the threading overhead would 
prevent any kind of speed-up after this mark...

Original comment by fal...@gmail.com on 25 Nov 2011 at 10:06

GoogleCodeExporter commented 9 years ago 
It's not that I actually use all those threads with numexpr. It just needlessly 
gives me error when everything is fine.

Original comment by xarthisi...@gmail.com on 25 Nov 2011 at 10:08

GoogleCodeExporter commented 9 years ago 
But I can also do some testing for >256 threads if you wish. Just tell me what 
you want to run.

Original comment by xarthisi...@gmail.com on 25 Nov 2011 at 10:10

GoogleCodeExporter commented 9 years ago 
Ah, I see your point now.  Yes, I'd raise the limit to 4096 (perhaps 16384).

And just out of curiosity, if you want to run "(sin(x)**2+cos(x)**2)" with x 
having 100 millions of doubles, and share the speed-ups in your big machine, 
that will be a lot of fun to look at the results :)

Thanks.

Original comment by fal...@gmail.com on 25 Nov 2011 at 10:25

GoogleCodeExporter commented 9 years ago 
Fixed in revision c432e60461a9.  Closing ticket.

Original comment by fal...@gmail.com on 25 Nov 2011 at 11:07