[Scip] Tweaking SCIP parameters for speed-up

[Sergey Smirnov]

Hi Michael,  

> You experienced a speed up factor of 2; on which test set or test
> instances, are these instances similar in their structure?

In all my tests I used sets of similar problems generated by AMPL. E.g. a set of ten traveling salesman problem stubs generated from the same model file but with different random seed.

> We are sure that parameter for a particular instance type can be improved,
> but how does your tuning works on an instances set like the MIPLIB?

Improving parameters for a particular instance type was the main use case for the tool when I was working on it. I think right now it is not very useful for the full MIPLIB at least because there are problems which hit limits with SCIP.

> Do you check the different results of the runs for feasibility/optimality
> or compare the solution value for each instance?

Yes, I compare the SCIP status and the primal value to the values from runs with default parameters.

> Is Turbo Boost and Hyper-threading disabled for your results and are the
> runs performed exclusively on a computer?

I run SCIP on virtual machines running under KVM hypervisor in a commercial cloud so I have no such control. The performance of a VM usually depends on the load made by other customers on the same hypervisor host. That is why I try to run SCIP with default parameters for every problem on every node. I can compare this time to other times I get on the same host.


Best regards,
Sergey




  
>  
> Best, Michael
>  
> > Hi Sergey,
> >  
> > I would be interested not so much in the implementation, but in the
> > methodological background.
> >  
> > Firstly it estimates how long it takes to solve single problems with
> > > default parameters.
> >  
> >  
> >  
> > How would you do that, in particular for those runs that hit the time
> > limit? I believe a LOT of people would be interested in such an estimate.
> >  
> >  
> >  
> > > Then we do a single run (for every problem) for every special value
> > > (e.g.
> > > -1 or 0) of every parameter and two runs for ranged parameters (bigger
> > > and
> > > lower than default values). That is about 3000 runs for every problem in
> > > a
> > > test set so I use a cluster of virtual machines in a cloud for the
> > > computations. After this step the tool compares the impact which
> > > parameters
> > > and their values make on the solving time.
> > >  
> >  
> >  
> >  
> > How? Considering a single parameter at a time? Or their combination? Or
> > pairs? Or ...
> >  
> >  
> > > Finally the tool takes four parameters (and their values) with the
> > > biggest
> > > impact and finds their best combination. This step is repeated for a fey
> > > times
> > >  
> >  
> >  
> >  
> > "This step"? Why repeating the four-parameter-combinations? For taking
> > into
> > account performance variability?
> >  
> >  
> > > and we get a settings file which speeds up computations noticeably for
> > > the
> > > given set of problems.
> > >  
> >  
> >  
> >  
> > Cool. Still would be interested in details.
> >  
> >  
> > Take care
> > Marco
> >  
> >  
> >  
> >  
> > --
> > Prof. Dr. Marco Lübbecke
> > RWTH Aachen University
> > Chair of Operations Research
> > Kackertstrasse 7
> > D-52072 Aachen
> > Germany
> >  
> > fon / fax: +49 241 80-93362 / 92369
> > marco.luebbecke at rwth-aachen.de (mailto:marco.luebbecke at rwth-aachen.de)
> > www.or.rwth-aachen.de/luebbecke (http://www.or.rwth-aachen.de/luebbecke)
> >  
> > <http://scholar.google.de/citations?user=cYIpR60AAAAJ>
> > <https://www.researchgate.net/profile/Marco_Luebbecke/>
> > <http://www.informatik.uni-trier.de/%7Eley/pers/hd/l/L=uuml=bbecke:Marco_E=.html>
> > <https://www.xing.com/profiles/Marco_Luebbecke>
> > <https://plus.google.com/u/0/100665281566037430285/>
> > <http://de.linkedin.com/pub/marco-l%C3%BCbbecke/7b/a2a/77/>
> > <https://www.facebook.com/OperationsResearchRWTH>
> > <https://twitter.com/#%21/mluebbecke> <http://klout.com/#/mluebbecke>
> > _______________________________________________
> > Scip mailing list
> > Scip at zib.de (mailto:Scip at zib.de)
> > http://listserv.zib.de/mailman/listinfo/scip
> >  
>  


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.zib.de/pipermail/scip/attachments/20140714/8a705153/attachment.html>