[SCIP] Is there any guidance reference of Multi-threaded calculation or SCIPsolveParallel(scip)?

[Ambros Gleixner]

Dear Melvin,

The usage of the concurrent solving mode is explained in

    https://scipopt.org/doc-7.0.1/html/CONCSCIP.php

To be able to call it, you need to configure your cmake with -DTPI=... / 
-DTPI_DIR=..., see

    https://scipopt.org/doc-7.0.1/html/CMAKE.php

(E.g. clone https://github.com/tinycthread and "cmake -DTPI=tny 
-DTPI=/path/to/tinycthread".)

Whether it helps performance will highly depend on your problem.  The 
general effectiveness of concurrent optimization is rather limited.


The UG framework is independent of the concurrent parallelization mode 
above.  Because it systematically parallelizes the branch-and-bound 
search, it generally holds more potential, but also here it will depend 
highly on your problem whether it can help.  The README in the UG 
directory is a starting point for how to set it up technically.

Best regards,
Ambros



Am 02.10.20 um 09:06 schrieb Melvin Shih:
> Dear SCIP team members,
> 
> I have built a model with SCIP in C++.
> Using example data, default SCIP can quickly obtain optimal solution in 
> minutes.
> Since the variables/constraints of the model increases, heavy 
> calculation effort is required and solving time obviously extends 
> without surprise.
> Therefore, I am studying branch-and-cut-and-price technical details 
> through SCIP Doxygen Documentation to create my rules.
> It seem not easy to overcome in short time.
> 
> On the other hand, I put my model onto a High-Performance Computing Service.
> It shows the usage of CPU is limited by C++.
> Multi-threaded calculation should be able to help.
> Also, ParaSCIP shows another option.
> However, I am not able to connect the ideas between std::thread and 
> ParaSCIP.
> SCIP_CALL(SCIPsolveParallel(scip) also has been tried and it doesn't 
> work.  (system shows: SCIP was compiled without task processing 
> interface. Parallel solve not possible.)
> Is there any example/evidence to include Multi-threaded calculation 
> based on my existing model?
> I wonder what I can do before having my own rules.
> 
> Sicerely,
> 
> Melvin
> 
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> https://listserv.zib.de/mailman/listinfo/scip
>