[SCIP] Setting relative gap in FiberSCIP and ParaSCIP
Stefan Vigerske
svigerske at gams.com
Wed May 3 08:30:45 CEST 2023
Hi,
Yuji had some quick look and it seems that there is no easy fix for this
situation. One reason could be that SCIP alone can do restarts, but for
SCIP in FiberSCIP, this is currently disabled. However, there seem to be
also other changes in SCIP in recent years that FiberSCIP hasn't caught
up with yet. We'll have to wait until FiberSCIP/ParaSCIP gets back to
the top of Yujis long todo list.
Stefan
On 30/04/2023 20:51, Vladimir V. Voloshinov wrote:
> Dear Yuji,
> I tried the option "-sl gap.set" you wrote about. It works and I see a
> non-zero gap in fscip log!
>
> BUT, there is the next very unpleasant problem: parallel, multithreaded,
> fscip *runs much slower than one-thread scip* !!??
> And, if I remember correctly, I saw the same effect for FiberSCIP of older
> versions too....
> For example, on my Dell comp, CPU i5-8250U 3.40GHz, 32Gb, with disabled
> Hyperthreading (to have "real" 4 CPUs) I run SCIP 8.0.3 and FSCIP 8.0.3 on
> the same problem [burma14
> <http://comopt.ifi.uni-heidelberg.de/software/TSPLIB95/tsp/burma14.tsp.gz>]
> from TSPLIB95 with limits/gap=1.e-01. Here SCIP had found a solution in
> ~800 sec and FSCIP (run on 4 CPUs, with "-sth 4") - in ~1400 sec.
>
> Please, can you tell me what is the possible reason for this strange
> situation?
>
> SCIP and FiberSCIP have been run by commands (on Linux)
> *$ scip -s scip.set -f tsp_burma14_1.cip -l
> tsp_burma14_q.scip803.dell.gap_0.1.log.txt*
> *$ fscip ug.set tsp_burma14_q.cip -s scip.set -sl gap.set -sth 4 -l
> tsp_burma14_q.fscip803.dell.gap_0.1.log.txt*
> Just in case, I attached tsp_burma14_q.cip, scip.set, gap.set and log files
> to the message.
> I should say that TSP has been intentionally represented as 0-1 Quadratic
> Programming problem, similar to Quadratic Assignment problem, with bilinear
> objective function and linear constraints.
>
> Best regards,
> Vladimir.
>
>
> On Mon, Apr 24, 2023 at 9:46 AM Yuji Shinano <shinano at zib.de> wrote:
>
>> Dear Vladimir,
>>
>> Stefan Vigerske and I tried to implement it so that it works more
>> precisely.
>> Probably, the code is already released in the version.
>> Please try to use the command line option with "-sl gap.set", and the file
>> gap.set inclueds
>> "limits/gap" or "limits/absgap". Then, it would work better than the
>> previous versions.
>> Please let me know the results.
>> @Stefan: If you can say something else, please add your comments.
>>
>> Best regards,
>> Yuji
>>
>> On 24.04.23 1:06, Vladimir V. Voloshinov wrote:
>>
>> Dear Yuji,
>> excuse for bothering you, but I see that SCIP Opt. Suite 8.0.3 includes
>> UG 1.0 already.
>> I see that the parameter CheckGapInLC, which we tried to use, has been
>> excluded from the list of UG options, so it seems that some changes with
>> "gap management" happened...
>> But I see that the gap set in the scip.set file does not affect FiberSCIP
>> which still solves problem till zero gap.
>> Moreover, I see in ug/src/ug_scip/scipParaSolver.cpp, lines #1259,#1260,
>> that "limits/gap" and "limits/absgap" are "hardcoded" to be 0...
>> Maybe we missed something, but is it possible to set a non zero gap for
>> FiberSCIP (and ParaSCIP) now, since version 8.0.3? It is very important in
>> some circumstances...
>>
>> Best regards,
>> Vladimir
>>
>> On Mon, Sep 12, 2022 at 5:36 PM Yuji Shinano <shinano at zib.de> wrote:
>>
>>> Hi Sergey,
>>>
>>> On 12.09.22 16:17, Sergey Smirnov wrote:
>>>> Thank you for checking my instance.
>>>>
>>>> We also tried setting CheckGapInLC=TRUE but it did not help either.
>>>
>>> Thanks for testing it, too. Probably, I tried to implement it once,
>>> but I had some issues realizing it. Before UG version 1.0,
>>> UG was a bit hard to specialize for a specific "base solver"
>>> parallelization.
>>> In current UG version 1.0, it is easy to implement some features
>>> specialized to a specific "base solver".
>>> I will let you know when I add the feature to FiberSCIP and ParaSCIP.
>>>
>>> Best,
>>> Yuji
>>>
>>>>
>>>>> On 11 Sep 2022, at 17:42, Yuji Shinano <shinano at zib.de> wrote:
>>>>>
>>>>> Hi Sergey,
>>>>>
>>>>> Thanks for the instance and configuration. I checked it.
>>>>> Sorry, in the current implementation of FiberSCIP and ParaSCIP,
>>> controlling it only with the SCIP parameter looks difficult when it takes
>>> time to solve
>>>>> (if it can be solved in the racing stage, the "NoPreprocessingInLC =
>>> TRUE" might work).
>>>>> And it looks like the setting SCIP parameters for that is not good for
>>> FiberSCIP and ParaSCIP in a sense how it work is not clear.
>>>>> I will add the feature rather UG side in the future. I think it can
>>> not control vividly, but when LC recognized that gap is reached,
>>>>> then LC interrupts all computations from the UG side.
>>>>>
>>>>> Best regards,
>>>>> Yuji
>>>>>
>>>>> --------------
>>>>> Hi Yuji,
>>>>>
>>>>> Here are the instance and configuration files attached.
>>>>>
>>>>> Thank you in advance,
>>>>> Best regards,
>>>>> Sergey
>>>>>
>>>>>
>>>>>> On 10 Sep 2022, at 08:55, Yuji Shinano <shinano at zib.de> wrote:
>>>>>>
>>>>>> Hi Sergey,
>>>>>>
>>>>>> Can you send me the instance and parameters (UG and SCIP)? I will
>>> check it.
>>>>>>
>>>>>> Best regards,
>>>>>> Yuji
>>>>>>
>>>>>> On 10.09.22 6:50, Sergey Smirnov wrote:
>>>>>>> Hi Yuji,
>>>>>>>
>>>>>>> Thank you for the suggestion. I tried to set NoPreprocessingInLC =
>>> TRUE in ug.set but both fscip and parascip still solved till zero gap.
>>>>>>>
>>>>>>> Best regards,
>>>>>>> Sergey
>>>>>>>
>>>>>>>> On 9 Sep 2022, at 21:20, Yuji Shinano <shinano at zib.de> wrote:
>>>>>>>>
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> This is one of the frequently asked questions. FiberSCIP/ParaSCIP
>>> solves the presolved instance in the Solvers.
>>>>>>>> Therefore, the value is applied to solve the presolved instance.
>>> If you would like to control the gap more precisely,
>>>>>>>> one way would be to set
>>>>>>>> NoPreprocessingInLC = TRUE
>>>>>>>> In the UG parameter file.
>>>>>>>> However, this means that layered presolving does not work and may
>>> decrease its performance.
>>>>>>>> Please try to run with the above parameter.
>>>>>>>>
>>>>>>>> Best,
>>>>>>>> Yuji
>>>>>>>>
>>>>>
>>>>>
>>>>> On 09.09.22 19:45, Sergey Smirnov wrote:
>>>>>> Hello, SCIP community,
>>>>>>
>>>>>> I have problems setting relative gap in parallel version of SCIP
>>> 8.0.1.
>>>>>>
>>>>>> When I set limits/gap = .5e-01 in scip.set and run scip like scip -f
>>> problem.cip -s scip.set scip stops after reaching the gap specified.
>>> However, fscip and parascip seem to ignore this parameter. For example, I
>>> run fscip the following way and it solves until optimality:
>>>>>> fscip ug.set problem.cip -s scip.set
>>>>>>
>>>>>> What is the correct way to set relative gap in fscip/parascip?
>>>>>>
>>>>>> Thank you in advance,
>>>>>> Best regards,
>>>>>> Sergey
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> Scip mailing list
>>>>>> Scip at zib.de
>>>>>> https://listserv.zib.de/mailman/listinfo/scip
>>>>>
>>>>>
>>>>> --
>>>>> --------------------------------------------------
>>>>> Yuji Shinano
>>>>> Zuse Institute Berlin
>>>>> Takustrasse 7, 14195 Berlin-Dahlem, Germany
>>>>> Phone:+49 30 84185-477, Fax:+49 30 84185-269
>>>>>
>>>>> _______________________________________________
>>>>> Scip mailing list
>>>>> Scip at zib.de
>>>>> https://listserv.zib.de/mailman/listinfo/scip
>>>>
>>>
>>>
>>> --
>>> --------------------------------------------------
>>> Yuji Shinano
>>> Zuse Institute Berlin
>>> Takustrasse 7, 14195 Berlin-Dahlem, Germany
>>> Phone:+49 30 84185-477, Fax:+49 30 84185-269
>>>
>>> _______________________________________________
>>> Scip mailing list
>>> Scip at zib.de
>>> https://listserv.zib.de/mailman/listinfo/scip
>>>
>>
>>
>> --
>> Vladimir V. Voloshinov,
>> Ph.D, head of lab. C-3 "Distributed computing algorithms",
>> http://www.iitp.ru/ru/researchlabs/1040.htm,
>> Center for Distributed Computing, Institute for Information Transmission
>> Problems RAS, http://www.iitp.ru
>> web: GoogleScholar profile
>> <https://scholar.google.ru/citations?hl=en&user=-m4QhNEAAAAJ&view_op=list_works&sortby=pubdate>
>>
>>
>>
>> --
>> --------------------------------------------------
>> Yuji Shinano
>> Zuse Institute Berlin
>> Takustrasse 7, 14195 Berlin-Dahlem, Germany
>> Phone:+49 30 84185-477, Fax:+49 30 84185-269
>>
>>
>
>
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> https://listserv.zib.de/mailman/listinfo/scip
More information about the Scip
mailing list