On Fri, Dec 14, 2012 at 7:24 AM, Konrad Hinsen
<[email protected]> wrote:
> Raul Miller writes:
>
>  > Even then your dominant cost is reading the data off of disk.  If the
>  > data is converted on the way in from the disk to RAM, and most of your
>  > CPU time is spent waiting for the disk, your conversion time might not
>  > even be measurable.
>
> My situation would be that the data is read to memory, ending up in a
> NumPy array, and then I have to copy it to a J array in a subsequent
> step. That's why I am interested in exploring the shared data approach.

Yes... I was thinking that this aspect might be redesigned.

>  > Note that if your data structure occupies half of your physical
>  > memory, J intermediate results might be expensive -- my rule of thumb
>  > (just a guess based on raw data size, until I get real measurements)
>  > is to expect 5x memory overhead from J.
>
> That's also a problem with NumPy, but there I know how to get around
> it.  With J, that remains to be explored.

Typically, I get my algorithms working with smaller data sets and then
introduce some kind of mechanism to break the content up in blocks.
Initially, I start with very small blocks (like 3 elements or
something) to make sure that I have my edge conditions set up
properly, and then I bump the blocks up to something useful (like
something in the 10s of megabytes).

Thanks,

-- 
Raul
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

Reply via email to