Use multiple machines with different algorithmic strategies in a race to solve an LP or MIP

The distributed concurrent solver uses a simple approach to take advantage of multiple machines. It starts an independent solve using a different strategy on each machine. The machines then race to see which can solve the model first. The solve is done when the first machine crosses the finish line. By trying different strategies at the same time, the concurrent optimizer can often find a solution faster than if it had to choose a single strategy. Performance testing demonstrates the benefits of the distributed concurrent solver. In our testing, using 2 machines produces mean improvements of 1.1X over a broad set of difficult LP models; using five machines produces mean improvements of 1.7X over a broad set of difficult MIP models.

When to use distributed concurrent

Distributed concurrent is particularly effective on models whose solve times vary substantially depending on the data used to define the model or the algorithm used to solve it. Trying different strategies on different machines increases robustness and can smooth out variations in solve times.