On Mon, Dec 17, 2018 at 10:59:59PM +0100, Olivier Borderies wrote:
On Mon, Dec 17, 2018 at 10:28 PM serge guelton <
serge.guelton@xxxxxxxxxxxxxxxxxxx> wrote:
On Wed, Dec 12, 2018 at 12:19:28PM +0100, Olivier Borderies wrote:
>
> I can reproduce your problems, and after a look at the parallel run,
tehre
> are *a lot* of shared access in the parallel loop, and you're not
> protecting them. All accesses to n_accept, redefinition of new_path
etc are
> done on shared variable...
>
>
> Thx for the pointer ! Indeed I'm new to omp. I fixed it. Now it does
work
on
> macOS and Linux. But surprisingly the speedup is much larger in macOS
(laptop
> and desktop) than on the Linux VM.
> On macOS the speed up from omp (branch omp) is very much in line with
that from
> concurrent.futures (branch master) - as expected since each top loop
runs
an
> entirely separate simulation.
> But on Linux, the speed up is far less. No crash, no warning, good
results
> though. Since it is a VM I cannot really see the CPU use (and check it
climbs
> to 100%). I checked that 'apt list --installed | grep omp' shows
'libgomp1'....
> What kind of reason could explain that ??
Inside a VM... how many core did you allocate to the VM?
The VM I picked had 16 cores but for unexplained reasons it would not go x~16
faster as it does on my laptop and desktop (x4 and x8 respectively, for that
number of cores). It's exactly the same code though.. Bizarre... Is there some
special characteristic that a core must have to work well with omp ?..
If you're ultra-motivated, you could reformat your README to fit into
http://serge-sans-paille.github.io/pythran-stories/
That would be a nice contribution to the project ;-)
I was planning to do a write-up in the form of a medium article + links. There
are many topics: Pythran (first and foremost), Simulated Annealing algo, TSP
problem, (permutation signature - not yet in code/repo), parallelism via
concurrent futures and OMP (at least it works as expected on my macOS
machines), remote Jupyter with cloud VM (Google and Azure).
So maybe the best way is to make the Pythran part + context into an .md for
you
blog, and have the medium article link to it. What do you think ?
The right template would be https://github.com/serge-sans-paille/
pythran-stories/blob/master/content/2018-11-28-prototype-to-fast-c%2B%2B.rst
for example, correct ?
