Answers to: Solving large problem with free LP solvers (50+ million variables)http://www.or-exchange.com/questions/9129/solving-large-problem-with-free-lp-solvers-50-million-variables<p>I was wondering if any of the free solvers were powerful enough to solve a problem that has approximately 50M variables and 100+M constraints. It's a maximization and is entirely linear.</p>
<p>I was thinking of using Coin-OR's CLP solver on a relatively powerful workstation.</p>
<p>Will this yield a solution? If not, what is the rule of thumb for a free LP solver's maximum problem size? Would a commercial solver (e.g. cplex or gurobi) handle such a problem?</p>
<p><em>Note: I realize the answer depends on the problem type and solver, but I just wanted to ensure that I'm not off by multiple orders of magnitude before I actually start. The largest problems I could find references to (on a Google search) were less than 10M variables.</em></p>enMon, 27 Jan 2014 09:20:29 -0500Answer by Erling_MOSEKhttp://www.or-exchange.com/questions/9129/solving-large-problem-with-free-lp-solvers-50-million-variables/9144<p>Most likely the dual problem will solve more efficiently than the primal since you have more rows than columns. Moreover, most likely an interior-point based algorithm works better than a simplex algorithm due to the size. The latter means that you most likely should use a commercial code. Since the public domain interior-point based codes for large scale LPs are not that good. [Please correct me if I wrong and biased.]</p>
<p>You will need a large computer in any case but how large will depend on the structure in your problem i.e. the sparsity pattern after presolve.</p>
<p>It would be fun if you would donate the problem to public domain so it can be used for test and benchmarking. I will be happy to try our code <a href="http://mosek.com">MOSEK</a> on it.</p>Erling_MOSEKMon, 27 Jan 2014 09:20:29 -0500http://www.or-exchange.com/questions/9129/solving-large-problem-with-free-lp-solvers-50-million-variables/9144