Presolve or >4 threads makes feasible non-convex model infeasible

I have a relatively small but non-convex quadratic optimization model that exhibits some strange behavior.

When I run it with more than 4 threads, Gurobi reports it is infeasible.

When I run it with Presolve not set to 0, Gurobi reports it is infeasible.

Otherwise (NonConvex=2 DualReductions=0 Presolve=0 Threads=4), Gurobi eventually finds the correct answer of just over 3.0.

The mps file is available here: https://rice.box.com/s/30c4t4vhyf9bw1wk40n4su1r99nuloxe

It's possible there are numeric issues because closely related models that don't exhibit this problem do give warnings like:

Warning: max constraint violation (1.8199e-04) exceeds tolerance

However, the coefficient range seems reasonable:

Optimize a model with 0 rows, 15 columns and 0 nonzeros
Model fingerprint: 0xc3aa9067
Model has 9 quadratic objective terms
Model has 16 quadratic constraints
Coefficient statistics:
  Matrix range     [0e+00, 0e+00]
  QMatrix range    [5e-01, 1e+00]
  QLMatrix range   [1e+00, 1e+02]
  Objective range  [0e+00, 0e+00]
  QObjective range [2e+00, 4e+00]
  Bounds range     [0e+00, 0e+00]
  RHS range        [0e+00, 0e+00]
  QRHS range       [4e+03, 5e+03]

Continuous model is non-convex -- solving as a MIP.

Presolve or >4 threads makes feasible non-convex model infeasible

Comments

Didn't find what you were looking for?