• Official comment
Gurobi Staff

Hello Manuel:

The issues you are seeing with both barrier and crossover restarts are largely due to the poor numerics, think very large ranges of the coefficients, for starters, on the order of 10^13, as can be seen from the log



Coefficient statistics:      Matrix range [1e-08, 1e+05]   Objective range [1e+00, 1e+00]             range [5e+01, 2e+06]         RHS range [3e-01, 4e+07]Warning: Model contains large matrix coefficient range



You may want to read through our guideline to numerical issues to get a better idea what typically work and what does not

https://www.gurobi.com/documentation/9.0/refman/num_grb_guidelines_for_num.html

Basically, when the barrier restarts (just as the crossover) we turn on more numerical "fail-safe" routines which in turn are more computationally expensive, but in some cases this does help to stabilize the computations.  Barrier restarts when the algorithm gets off-track or stagnates, and likewise, for the crossover phase; for the latter, you can also see that there are some changes to the basis being introduced, as most likely the starting basis candidate is nearly singular.

As you mention, rescaling the model helps, and this is indeed the way to go.

Besides the above link, if you are curios about the crossover itself, the below is a cool paper to read that should tell you a bit more about how crossover works,

http://theory.stanford.edu/~megiddo/pdf/bases.pdf

Hope this helps.

Dear Yuriy,

Thanks for your feedback.
I get the same warning and also wanted to know what to do to improve the situation.

My question is: What do you mean with "rescaling the model helps"?

Kind regards,
Markus

• Gurobi Staff

Hi Markus,

have you had a chance to review the guide to numerics

https://www.gurobi.com/documentation/9.0/refman/num_grb_guidelines_for_num.html

for example