I am using CPLEX for solving huge optimization models (more than 100k variables) now I'd like to see if I can find an open source alternative, I solve mixed integer problems (MILP) and CPLEX works great but it is very expensive if we want to scale so I really need to find an alternative or start writing our own ad-hoc optimization library (which will be painful)
Any suggestion/insight would be much appreciated
I personally found GLPK better (i.e. faster) than LP_SOLVE. It supports various file formats, and a further advantage is its library interface, which allows smooth integration with your application.
Have you tried lp_solve? There were also some other suggestions in the following questions, for Java:
I recommend checking out the COIN project. COIN OR
Many good solvers here, including ipOPT for nonlinear problems and a couple mixed integer solvers as well.
I am surprised that nobody has mentioned MIPCL (http://www.mipcl-cpp.appspot.com/index.html). This solver claims to be open source under LGPL license (source: http://www.mipcl-cpp.appspot.com/licence.html), thus it is also suitable to be used in non-open-source applications. But what is missing to be really open source is the source code of the solver itself.
Hans Mittelmann very recently (10 Sep 2017) did Mixed Integer Linear Programming Benchmark: http://plato.asu.edu/ftp/milpc.html (you might also be interested in looking at the overview page http://plato.asu.edu/bench.html or the slides of his talk: http://plato.asu.edu/talks/informs2017.pdf).
In the Mixed Integer Linear Programming Benchmark with 12 threads and a time limit of 2 hours MIPCL managed to solve 79 instances. Only the commercial solvers CPLEX, Gurobi and XPRESS managed to solve more under the given constraints (86 or 87 instances, respectively).
Also in terms of the chosen performance metric (again using 12 threads) MIPCL is faster than the benchmarked SCIP derivates (FSCIPC, FSCIPS) and the open source solver CBC. Again only the commercial solvers CPLEX, Gurobi and XPRESS outcompete MIPCL significantly in terms of performance.
I'll add the following to the already excellent answers.
Whilst, as others have pointed out, commercial solvers are much faster and more capable than the free alternatives it's important to consider how much of an optimality gap you can accept. For large problems with many integer variables you may get much faster solve-times if you can accept 1% or even greater optimality gap (defaults tend to be around 0.01% or less).
Of course if you are solving a problem where 1% translates into millions of dollars this is not acceptable - hence the market for top-notch solvers. (Some comparisons of Gurobi to free solvers here)
I would agree with others that using a platform that generates solver-independent problem files (such as *.mps, *.lp files) is worthwhile as you can then try out other solvers. I'm using PuLP and am finding it, and the free COIN_CBC solver, excellent. Although limited for problems with many integer variables.
Not open source, but if you have a Microsoft Academic Alliance license, the Microsoft Solver Foundation (MSF) enterprise edition is included. Gurobi is also free for academic purposes, I've used it in my thesis research.
