Use multiple machines that work together to solve a MIP model

The distributed MIP solver divides the work of solving a single MIP model among multiple machines. A manager machine sends different portions of the MIP search tree to each worker machine to solve, and it periodically rebalances the tree to put useful load on all the workers. Performance testing demonstrates the benefits of distributed MIP. For models in the MIPLIB 2010 benchmark set that take more than 100 seconds to solve, distributed MIP on 8 machines produces a mean improvement of 2.87X. Distributed MIP can also dramatically reduce solve times on some of the most difficult models. For example, model seymour, a set covering model in MIPLIB 2010, takes 9,267 seconds to solve on a single machine. Distributed MIP on 32 machines reduces the solve time to 633 seconds — a 14.6x speed-up.


When to use distributed MIP

Distributed MIP works best on difficult models with large search trees. These models have enough work to ensure all the workers stay busy. Easier models that solve at the root won’t benefit from the extra workers.