[Coin-discuss] Contribution ?

[Bertrand Le cun]

Hi,

>     > 1- Why use PVM or Serial communication library, MPI, or Multithreaded
>     > libraries like PM2 (http://www.pm2.org/), or other could be a better choice. 
>     > An application could be executed on Cluster of SMP, (like IBM SP3, i think).
>     
>     The reason is that at the time the project started MPI-2 implementations
>     weren't around at all, therefore spawning new processes was impossible. And we
>     really wanted the ability of deleting processors or adding new processors to
>     the configuration on the fly thus harnessing unused workstations. Also, for
>     the purposes of branch and cut and price (unlike for branch and bound) latency
>     and bandwith is not a big problem since search tree nodes are processed for a
>     relatively lengthy time.
Yes and no.
There are some works that has shown, that the biggest time part of the resolution
in Branch and Price/Cut is on the Evaluation of the Root Node. And finally, the 
conclusion is : to obtain performance or "speed-up", you have to parallelize also 
the evaluation function. 
It was the reason why when i decide to develop Bob++, I would to design a kind 
of kernel on which several algorithms (Branch and Bound, Dynamic Programming) could
be executed. This kernel decides to execute an algorithm in parallel or not.

I am surprised that you want dynamicity for process creation.
Does it means that you want to write a "network analyser" to decide which machines
could be involved in the virtual machine ?
In most of case, the application is started on a fixed number of processor, and after
the "parallel" algorithm has to decide which processor executes the operation on a 
Branch and X" node.
It seems that most of the parallel communication libraries do not support dynamic process
creation, due to parallel computer companies (like IBM, i guess), because they want
processor reservation system. 

>     True, at the moment it is geared towards problems that have an LP relaxation.
>     It could concievably extended to further generalize the solver to be a generic
>     lower bounding technique.
One more abstract Branch and Bound class 

>     Adding lp_solve or soplex interfaces would be very good, we just lack the
>     manpower to do it. If anyone is interested in contributing... :-) It should
>     not be too hard.
I am here ;-).
>     
>     If there is enough interest, we can make nightly
>     tar files out of it.
CVS is OK, i will take it. 
>     
>     We are definitely interested and you found the right place to ask :-).
>     ....
>     source projects) is a brutal meritocracy :-).
Very fine. 
I take the development sources in the CVS tree, i study it. And i will recontact you...


Bertrand-
+----------------------------------------------------------------------+
|                           Bertrand LE CUN                            |
+----------------------------------------------------------------------+
| Universite de Versailles St Quentin, Laboratoire PRiSM, OPALE Team   |
| 45, avenue des Etats-Unis 78000 Versailles                           |
| email: Bertrand.Lecun at prism.uvsq.fr       Phone: (+33 1) 39 25 40 50 |
| Web  : http://www.prism.uvsq.fr/~blec
+----------------------------------------------------------------------+