Solve Even Faster
Gurobi Distributed Optimization lets you leverage multiple machines to reduce solve times. Some optimization models solve 15 times faster with 32 machines, and speed-ups of 2-3x are common with eight machines. Gurobi offers three distributed algorithms:
- Distributed MIP — where multiple machines work together to solve a single MIP model
- Distributed concurrent — where multiple machine use different algorithmic strategies in a race to solve an LP or MIP model
- Distributed tuning — where multiple machines do experimental solves to find parameter settings that improve performance
All three distributed algorithms are easy to use. Simply install Gurobi Remote Services on the set of machines that you wish to use as distributed workers, then use the same Gurobi commands that you would use if you were going to perform optimization or tuning on just a single machine. A Gurobi parameter allows you to specify the names of the worker machines that are available for running distributed algorithms, and another parameter indicates how many jobs to start on these workers. The Gurobi Optimizer handles all the work of dividing the computation among the machines. The speed-up achieved by distributed optimization varies depending on the model.
How it Works
Machine Requirements
The distributed algorithms require a set of machines to serve as distributed workers. These algorithms work best on identical machines, but it is fine if the machines vary slightly. Distributed MIP works best on 8-32 machines, distributed concurrent is most effective on 2-16 machines, and distributed tuning can take advantage of all available machines.
Licensing
You can use any of the distributed algorithms by adding the distributed capability to an existing named user, single machine, or compute server license. A license is only required for the manager machine that launches and coordinates workers. Worker machines do not need separate licenses; they only need Gurobi Remote Services installed.
Know before you buy
Do you want to know if your models can benefit from distributed optimization? Simply send us your models as MPS or LP files. We will run them on a cluster of machines and tell you if:
- Distributed MIP solves your models faster
- Distributed concurrent decreases variations in solve times
- Distributed tuning finds parameter settings that speed up single machine solve
