Benchmark

It has come to our attention that a fair number of people have been trying to use the NEOS Server to benchmark their applications across several solvers. Of course, by the distributed nature of our system, often times those problems are run on very different machines so that the solvers cannot easily be compared. This benchmarking solver allows multiple solvers to tackle a job on the same machine sequentially.

We also offer an independent analysis of the solutions returned by the solvers as well as a look at the model as given to the solvers.

Using the BENCHMARK solver on NEOS

This benchmark solver requires an AMPL model of the problem. Problem data may be submitted as a separate file. A commands file is allowed and will be run once for each solver. The solver will be set externally and should not be set in the commands file. Solver options for the various solvers chosen may be set here. These will override the solver options we have chosen as defaults, which generally minimize output and sometimes allow for larger problems than the solver's original defaults. If no solve command exists, it will be added. If a solve command does exist, it should be on its own line. Please do not use the word solve elsewhere in the commands file. A command to print timing information on the solve will be added to the commands file.

This solver also offers two types of analysis. The problem analysis prints out various characteristics of your problem (e.g., the number of nonlinear variables in the constraints). If the problem analysis conflicts with your knowledge of the model, it may be that AMPL is reducing the problem in its presolve phase, which can be turned off with the presolve option. The problem analysis we give summarizes the model as it is received by the solver.

Our solution analysis attempts to show whether the solution returned by the solver is in fact a reasonable solution to the problem. You decide what is reasonable. In particular, it is helpful if you give us limits on the size of the complementarity error and scaled optimality error (scaled by dividing by the 2 norm of the gradient at the solution). There is a certain trade-off that can be made between the two, depending on the tolerance we use for deciding whether a constraint is active at the solution. Our analysis will try up to three activity tolerances to try to find a complementarity error and scaled optimality error pair that both fall below the limits you give. We also return the feasibility error of the solution and an idea of the error in our analysis. If that error is not low enough, then the complementarity and optimality errors of the solver may actually be better than we show. For more information check either the [PDF] or [postscript] version of our explanation of the solution analysis. Enter the location of the ampl model (local file)
Model File:

Enter the location of the ampl data file (local file)
Data File:

Enter the location of the ampl commands file (local file)
Commands File:

Accepted Complementarity Error:

Scaled Optimality Error:

Acceptable Feasibility Error:

filter:

IPOPT:

KNITRO:

LANCELOT:

LOQO:

MINOS:

PENNON:

SNOPT:

Independent analysis of solution:

Comments:

e-mail address:

Please do not click the 'Submit to NEOS' button more than once.

_{DOE disclaimer

DOE Web privacy policy}