Re: [Parralelisms]

Le 05/03/2014 18:00, Felix a écrit :
> it might be tricky to find a good partition, and your tree-view might
> give the right ideas. to me it does not make sense to schedule
> everything in advance. for example, you may want to start with just 4
> threads, and take the first of T1..T4 that finishes to become T5. you
> cannot know which, as it depends on which nodes have changed
> (a.is_changed(mm)).
>
> if there is a matrix where no such partition induces a useful
> parallelization scheme, i'd be interested to know!

Thank you for the answer, you are right, the dependance tree might be 
not the good way for implementation. I think i had in mind massively 
parallelism instead of simple parralelism.
I will try to implement in openmp a more simple solution.

Le 08/03/2014 18:00, [email protected] a écrit :
> I think you are making it harder than it needs to be and
> overlooking why/where it could be of benefit.
>
> That sample matrix is not in the form you actually get.
>
> The form you get is bordered-block-diagonal, possibly
> hierarchical.  By hierarchical I mean you will find a BBD form
> inside a block.
Ok i will change my example to do test.
> Blocks can be done in parallel with other blocks, then the
>
> In the netlist, blocks can be identified by subcircuit/module
> instantiations.  Borders can be identified by the next higher
> level and subcircuit/module calls.
>
> The netlist node list is an adjacency matrix.
>
> This stuff is available.  It will be faster and easier to use it
> rather than to try to extract that from the matrix.  By faster,
> I mean both faster to run and faster to develop.
>
> Any block smaller that about 100 nodes is not worth decomposing
> further.  Just assume it is connected.
I will try to check , if it necessary . Futhermore i think we don't more 
threads than the number of threads that can run simultaneously(num_procs 
* hyper_threading factor ) so most probably a decomposition into 16 
threads will be enough for the current processor architecture.
> If you have a block that is big enough to split, the way to
> split is to look for ways to move nodes to a border.  This is
> the opposite of what traditional ordering methods do.
>
> The block requirement for parallel is the same as for
> incremental update.
>
> Global ordering takes at best O(n log(n)) time.  This is too
> slow.  The way to make it faster is local ordering and reuse.
>
> Gnucap's ordering is non-optimal, but needs to stay early in the
> process.  It should be done before expansion.  That is one
> reason I never did anything with the post-expand ordering hook.
>
And sorry for the late answer , i have administrative stuff to do at 
work that i always pospone. I come to you to let you know how i progress.

Thanks fo your reply.

-- 
Beranger Six,
CCamy système
06 33 16 10 17