Re: Multicore problem

[jonathan <johnscpg@googlemail.com>]

On Mar 9, 6:48 pm, Anatoly Chernyshev <achernys...@gmail.com> wrote:
> > Perhaps the two tasks/cores are competing for data-cache space, and
> > causing more cache misses than if a single task/core is active?
>
> > How large is your 3D array, compared to the cache?
>
> Right now it is small, 23 Mb, for the purposes of algorithm honing.
> 2Mb L2 cache. In perspective the array could be 4-8 Gb (not for my
> current machine, of course).
>
> > If this is the problem, you could try to divide the whole problem into
> > slices so that each slice processes a part of the 3D array that *can*
> > fit in the cache, but can also be processed in parallel by two
> > tasks/cores. Not knowing what your processing actually does I can't say
> > more about how to slice the problem.
>
> Each task begins in the cell of its designated area of array, and then
> does a Markov walk all around the place using the cell pointers.

Of course I can't say much about your specific problem.  I can tell
you
(the little) I learned battling something similar (using linux and
GNAT).  The problem I was up against was latency: minimum time to send
data, no matter how small the data, between 2 cores on the same
processor.
I measured this time as about 40 microseconds, (exchange of data
between
cores). Might have been 20 or 60, but think 10's of microseconds.
If message passing (or sharing data in an array) is more frequent
than this, you'll never see a speed up on the several cores.
(Some expensive networks advertise
latencies of a few microseconds, and some experimental operating
systems advertise sub-microsecond latencies.) So you should not
think of passing data more frequently than say every 200 or 400
microseconds if you want to see speedup.  Which is a long
time .. you can do 10's of thousands of floating point
operations in that time!

Jonathan