DO NOT MERGE! Benchmarks: Commercial solvers run

[KristijanFaust-OET]

Implements CPLEX, knitro and xpress commercial solver runs.

Solver options

Knitro

Knitro support confirmed the following options

feastol=1e-6
opttol=1e-6
mip_opt_gap_rel=1e-4

• how to say we don't need a basic solution?

Xpress

• Have we heard back from FICO support?

Gurobi

Gurobi team recommended

      MIPGap: 1.e-4 # default
      FeasibilityTol: 1.e-6 # default
      OptimalityTol: 1.e-6  # default
      SolutionTarget: 1  # basic not needed

They don't think it's worth trying to set BarConvTol as the meaning of this varies too much between solvers.

Cplex

Cplex support confirmed

            "mip.tolerances.mipgap": mip_gap,
            # LP tolerances
            "simplex.tolerances.feasibility": lp_tol,
            "simplex.tolerances.optimality": lp_tol,
            "solutiontype": 2, # Basic solutions not needed

• Have asked: do the simplex tolerances also work for barrier?

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
solver-benchmark	Ready	Preview, Comment	Apr 15, 2026 5:43pm

[danielelerede-oet]

@KristijanFaust-OET @siddharth-krishna I've added Mosek to the set of commercial solvers again considering default options only.

In addition, I see you're using solver_model.get_time() for CPLEX, but that metric does not provide a clean solver runtime comparable to other solvers. It reflects an internal clock rather than the time spent solving a single instance, and may include additional overhead or accumulated time. As far as I know, CPLEX does not expose a directly comparable runtime metric through this interface. There is some ongoing work in Linopy to improve this (PyPSA/linopy#583) while I also provided a workaround that we could use here (PyPSA/linopy#636).