Solver Options for EPM

Contents

Solver Options for EPM#

EPM solves large-scale mixed-integer programs (MIPs). Properly configuring solver options can significantly reduce runtime and improve numerical stability.

Solver settings are defined in a cplex.opt file.
→ See example cplex.opt for reference.

Recommended Solver Parameters#

Parameter	Recommended Value	Description & Rationale
`threads`	`8–12`	Number of solver threads. Using more threads than CPU cores (oversubscription) can improve performance through load balancing and OS scheduling. Too many threads (e.g., >16) may increase overhead and slow the model. Test different values.
`lpmethod`	`4`	Controls which algorithm is used to solve LPs: `0`: Primal simplex `1`: Dual simplex `2`: Network `3`: Barrier `4`: Automatic (recommended) Let CPLEX choose the best method per instance.
`startalg`	`4`	Starting algorithm for root LPs (when solving MIPs): same values as `lpmethod`. Default `4` allows CPLEX to choose the best approach dynamically.
`predual`	`-1`	Tells CPLEX to solve the primal (`0`), dual (`1`), or to choose automatically (`-1`). Impact is model-specific; automatic is generally safe.
`baralg`	`1`	Barrier algorithm type: `0`: Standard barrier `1`: Barrier with crossover (recommended) Produces a basic solution usable by downstream steps.
`barcrossalg`	`1`	Algorithm used for crossover after barrier: `0`: Primal simplex `1`: Dual simplex (recommended)
`barcolnz`	`500`	Sets threshold for treating columns as dense in the barrier method. This can greatly impact performance. Try tuning in steps of ±100.
`barepcomp`	`1e-5`	Complementarity tolerance for barrier method. Lower values improve precision but may slow convergence. `1e-5` offers a good trade-off.
`solutiontype`	`2`	Type of solution to return: `1`: Basic solution `2`: Interior point solution. Use with caution on MIPs—primarily useful for LPs. (`solutiontype = 1` may be slower, but also more useful in post-analysis like Monte Carlo where specific properties for the solution are a nice-to-have).
`memoryemphasis`	`1`	Emphasizes memory-saving strategies, helpful for large models prone to memory issues.
`bardisplay`	`2`	Controls verbosity during barrier solve. Higher values produce more output (good for diagnostics).
`scaind`	`1`	Controls constraint matrix scaling: `-1`, `0`: No scaling `1`: Automatic scaling (recommended)
`auxrootthreads`	`0`	Disables auxiliary tasks in the root node. Reduces time spent in presolve when root node is slow to launch.

Notes on Performance Tuning#

Start with default recommendations and test on a small-to-medium scenario.
Barrier method parameters (baralg, barcolnz, barepcomp) can greatly impact performance, especially for LP relaxations.
Interior point solution (solutiontype = 2) is not always suitable for MIPs. Use only when appropriate.
If presolve seems to hang, try setting:
```
auxrootthreads = 0
```

Resources#