Solving a mixed integer programming (MIP) problem can be extremely time-consuming using the so-called brand and bound algorithm. Therefore, a MIP solver like MOSEK incorporates a lot of algorithmic improvements to reduce the solution time. Sometimes those improvements are called tricks.
Now to evaluate whether some trick benefits the MIP solver, then the MIP solver with and without the trick included is used to solve a benchmark set of test problems and if the benchmark results indicate the trick helps, then it is included in the solver.
Clearly, if all the test problems are solved on one specific computer, then the timing results are comparable. However, to make robust conclusions then a lot of carefully selected test problems must be employed. This has the unfortunate consequence that evaluating a new trick is very time-consuming. The benchmark problems can of course be solved in parallel, but solving multiple problems in parallel on one computer will not produce reliable timing results that can be compared. The only way of getting comparable timing results quickly is to solve many problems in parallel on a cluster of identical computers.
That is why we at MOSEK recently invested 55K+ USD in a compute cluster made up of identical computers for the MIP development team. The cluster consists of 4 boxes each containing 8 computational nodes i.e. it provides 32 identical computers.
Hopefully, you won't have to wait too long to see the benefits in Mosek arising from faster testing of potential new improvements in the MIP solver!
In the meantime, you can enjoy this photo of our cluster working away :)
