Originally posted by: A.Florio

I have code that works on Cplex 12.7 but has stopped working on 12.8. It is a simple TSP solver with SECs added dynamically.

Within the UserCutCallback, when the first SEC violation is identified (when solving the relaxation at the root node of branch-and-bound), I create IloRange and, when setting up the cut with setLinearCoefs, the execution simply halts. No exception is raised. CPU becomes idle.

Architecture: macOS 10.13.4

Apple LLVM version 9.1.0 (clang-902.0.39.1)

Code:

 

ILOUSERCUTCALLBACK2(TSPSolverMinCut, IloNumVarArray, xijs, Instance, instance) {

    const double CUTTOL=1e-2;

    IloNumArray valscplex(getEnv());

    getValues(valscplex, xijs);

    vector<vector<double>> vals(instance.numNodes()+1,

            vector<double>(instance.numNodes()+1, 0));

    for (int k=0, i=0; i<=instance.numNodes(); ++i)

        for (int j=0; j<=instance.numNodes(); ++j, ++k)

            if (i!=j)

                vals[i][j]=valscplex[k];

    valscplex.end();

    // subtour elimination constraints (fractional) (global cuts)

    {

        MinCutSolver mincut(vals);

        MinCutSolution sol=mincut.solve();

        if (sol.value()<1-CUTTOL) {

            IloRange sec(getEnv(), 1, IloInfinity);

            IloNumVarArray vars(getEnv());

            IloNumArray coeffs(getEnv());

            for (int i=0; i<=instance.numNodes(); ++i) {

                if (sol.minCut()[i]) {      // i is in the cut

                    for (int j=0; j<=instance.numNodes(); ++j) {

                        if (i!=j && !sol.minCut()[j]) { // j is NOT in the cut

                            vars.add(xijs[i*(instance.numNodes()+1)+j]);

                            coeffs.add(1.0);

                        }

                    }

                }

            }

            try {

                sec.setLinearCoefs(vars, coeffs);        // Execution is halting HERE

            } catch (...) {        // No exceptions raised

                cout<<"exception caught!"<<endl;

                exit(-1);

            }

            add(sec);

            vars.end();

            coeffs.end();

        }

    }

}

#CPLEXOptimizers
#DecisionOptimization

That seems familiar :-(

Can you please try the solution described here. Note that despite my claim, the fix did not make it into 12.8 :-(

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: A.Florio

I added env.setNormalizer(false) just before cplex.solve(), and also getEnv().setNormalizer(false) in the 1st line of the UserCutCallback, but the problem persists.

 

#CPLEXOptimizers
#DecisionOptimization

OK, I have to look deeper then. This may take a while.

In the meantime, do things work any better if you create 'sec' like so:

IloRange sec(getEnv(), 1, IloScalProd(vars, coeffs), IloInfinity)

? (after populating vars and coeffs)

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: A.Florio

Yes, great workaround, that solved the problem. Thank you very much for your prompt answers!

Also, setNormalizer(false) is not needed when using this workaround.

 

#CPLEXOptimizers
#DecisionOptimization

Thank you for your feedback!

I confirmed that the issue boils down to the same bug that is described in the thread I linked to.

I also confirm that in this variant of the problem, setNormalizer() is neither helpful nor needed to work around the problem.

While looking at your code I noticed something: there may be a memory leak since you are not end()ing 'sec'. If you look at the user cut examples iloadmipex5.cpp and ilobendersatsp.cpp you can see that a cut that is instantiated in the callback is always added like this

add(cut).end();

This is independent of the deadlock but may still improve your code.

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: A.Florio

Thanks, I did not know that!

 

#CPLEXOptimizers
#DecisionOptimization