Improving Barrier Performance

[Andrew B. Martin]

Hello,

I have a linear model with 889464 rows, 9855786 columns and 400848336 nonzeros that gets presolved down to 587751 rows, 7736078 columns, 68878492 nonzeros.

The model is an assignment problem with a transportation network. It is about 85% GUB constraints, and the other 15% describe the transportation network and restrictions on assignment.

When I try to solve the model using Dual Simplex, Gurobi reaches a point where the algorithm stops making progress. 

Barrier performs better, but the crossover phase is very slow (see below for an example from the log).

Sometimes the model hits numerical issues, so I'm running Gurobi with parameters BarHomogenous: 1, and NumericFocus: 3.

Do you have any recommendations about how I might get better performance from Gurobi for this model?

Thanks!

--Andrew

 5589348 PPushes remaining with PInf 3.5798720e+08              5791s

 5587623 PPushes remaining with PInf 3.5789888e+08              5796s

 5585754 PPushes remaining with PInf 3.5779663e+08              5801s

 5584104 PPushes remaining with PInf 3.5764103e+08              5806s

 5582004 PPushes remaining with PInf 3.5747465e+08              5811s

 5580414 PPushes remaining with PInf 3.5733484e+08              5816s

 5578839 PPushes remaining with PInf 3.5718552e+08              5821s

 5577299 PPushes remaining with PInf 3.5671755e+08              5826s

 5575503 PPushes remaining with PInf 3.5651111e+08              5831s

 5573758 PPushes remaining with PInf 3.5640113e+08              5836s

 5572077 PPushes remaining with PInf 3.5633012e+08              5841s

 5570259 PPushes remaining with PInf 3.5620209e+08              5846s

 5568525 PPushes remaining with PInf 3.5611027e+08              5851s

 5564955 PPushes remaining with PInf 3.5593708e+08              5855s

 5560028 PPushes remaining with PInf 3.5570904e+08              5861s

 5555623 PPushes remaining with PInf 3.5553032e+08              5866s

 5550242 PPushes remaining with PInf 3.5542669e+08              5871s

 5545117 PPushes remaining with PInf 3.5529824e+08              5876s

 5541118 PPushes remaining with PInf 3.5516995e+08              5880s

 5536450 PPushes remaining with PInf 3.5480876e+08              5885s

 5531290 PPushes remaining with PInf 3.5456321e+08              5891s

 5527580 PPushes remaining with PInf 3.5424455e+08              5896s

 5525285 PPushes remaining with PInf 3.5362318e+08              5900s

 5522883 PPushes remaining with PInf 3.5347307e+08              5905s

 5520405 PPushes remaining with PInf 3.5291165e+08              5911s

 5518654 PPushes remaining with PInf 3.5177199e+08              5916s

 5517214 PPushes remaining with PInf 3.5064184e+08              5921s

 5515545 PPushes remaining with PInf 3.5019612e+08              5926s

 5513766 PPushes remaining with PInf 3.4971738e+08              5931s

 5512076 PPushes remaining with PInf 3.4927685e+08              5936s

 5510324 PPushes remaining with PInf 3.4886709e+08              5941s

 5508690 PPushes remaining with PInf 3.4850958e+08              5946s

 5507139 PPushes remaining with PInf 3.4804120e+08              5951s

 5505470 PPushes remaining with PInf 3.4752530e+08              5956s

 5503857 PPushes remaining with PInf 3.4699653e+08              5961s

 5502479 PPushes remaining with PInf 3.4619623e+08              5966s

 5501045 PPushes remaining with PInf 3.4559497e+08              5971s

 5499590 PPushes remaining with PInf 3.4516097e+08              5976s

 5497984 PPushes remaining with PInf 3.4448592e+08              5981s

[Sonja Mars]

Hi Andrew,

Do you have a full log file that you can share?

I would try BarHomogeneous at default and NumericFocus set to 1 or 2. Setting NumericFocus to 3 tells Gurobi to be very, very careful with numerics and it can lead to slower performance.

Best regards,
Sonja


----
Dr. Sonja Mars
Gurobi Optimization - Technical Support

[Andrew B. Martin]

Thanks, Sonja.

I've temporarily switched to a smaller, more tractable, model. I'll follow up on this with a full log when I come back to the larger model.