[SCIP] Bliss causing free() crashes

[Christopher Hojny]

Dear Alexander,

the good news: you did nothing wrong. Setting misc/usesymmetry to 1 or 2
activates symmetry handling in SCIP, and thus, the computation of
symmetries. The only difference is the used symmetry handling method (1
for inequalities, 2 for orbital fixing). Moreover, you can control the
point of time at which symmetries are computed, e.g., by

presolving/symbreak/addconsstiming
propagating/orbitalfixing/symcomptiming

in the first and last case, respectively. However, since your problem
contains exclusively general integer variables, neither method will be
able to handle the symmetries in your problem, because we currently only
support symmetry handling on binary variables.

I am not totally sure where the error comes from, but I guess it is a
memory issue arising since the symmetry group is too large. Did you try
to trace back in which routine the crash actually arises? If it is
really the size of the symmetry group, you may consider the bliss patch
provided via

https://scip.zib.de/index.php#download

The patch allows to limit the number of generators of the symmetry group
that are computed. By default, we use a limit of 1500 generators, but
you can modify the parameter by changing

presolving/symmetry/maxgenerators

I hope this helps, but I can also offer to have a look into your
instance to figure out where the crash actually comes from.

Best,
Christopher



On 07.02.19 23:51, Alex Meiburg wrote:
> Hi all,
> 
> I build and installed SCIP through cmake with SYM=bliss, and I installed
> bliss libbliss-dev from Ubunutu's package repository, which uses version
> 0.73. The building found Bliss without any issue. After playing around it
> seemed that actually get it to use this featuer, I needed to "set misc
> usesymmetry 1", and then it crashes during presolve.
> 
> To be precise, my full sequence of actions is:
> $ scip
> SCIP> read problem_16.lp
> SCIP> set misc usesymmetry 1
> SCIP> presolve
> 
> and then it gives this output:
> 
> LP Solver <SoPlex 4.0.1>: barrier convergence tolerance cannot be set --
> tolerance of SCIP and LP solver may differ
> LP Solver <SoPlex 4.0.1>: fastmip setting not available -- SCIP parameter
> has no effect
> LP Solver <SoPlex 4.0.1>: number of threads settings not available -- SCIP
> parameter has no effect
> transformed problem has 1540 variables (0 bin, 1540 int, 0 impl, 0 cont)
> and 31462 constraints
>   31462 constraints of type <linear>
> 
> original problem has 92905 active (0.191749%) nonzeros and 92905
> (0.191749%) check nonzeros
> 
> presolving:
> (round 1, fast)       505 del vars, 3630 del conss, 0 add conss, 3585 chg
> bounds, 0 chg sides, 0 chg coeffs, 0 upgd conss, 0 impls, 0 clqs
> (round 2, fast)       910 del vars, 25608 del conss, 0 add conss, 3990 chg
> bounds, 0 chg sides, 0 chg coeffs, 0 upgd conss, 0 impls, 0 clqs
> (round 3, fast)       910 del vars, 28222 del conss, 0 add conss, 3990 chg
> bounds, 0 chg sides, 0 chg coeffs, 0 upgd conss, 0 impls, 0 clqs
> (round 4, exhaustive) 910 del vars, 28267 del conss, 0 add conss, 3990 chg
> bounds, 0 chg sides, 0 chg coeffs, 0 upgd conss, 0 impls, 0 clqs
> (round 5, exhaustive) 910 del vars, 28267 del conss, 0 add conss, 3990 chg
> bounds, 0 chg sides, 0 chg coeffs, 3150 upgd conss, 0 impls, 0 clqs
>    (0.2s) probing: 51/630 (8.1%) - 0 fixings, 0 aggregations, 0
> implications, 0 bound changes
>    (0.2s) probing aborted: 50/50 successive totally useless probings
>    (0.2s) symmetry computation started: requiring (bin +, int +, cont +),
> (fixed: bin -, int -, cont -)
> free(): invalid pointer
> Aborted (core dumped)
> 
> Is something wrong with my bliss installation? Am I misusing the
> misc/symmetry settings? When I tried building bliss myself, it wasn't
> creating a .so (only .a), so is there a build script that handles it? Once
> I have it working, I would like to compile the SCIP-SDP plugin in too,
> which means I'd like to not use a pre-built version. Thank you all!
> 
> -- Alexander Meiburg
> 
> 
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> https://listserv.zib.de/mailman/listinfo/scip
>