Parallel Simulated Annealing with MRAnneal

Simulated annealing algorithms, which repeatedly make small changes to candidate solutions to find approximately optimal ones, are a common method for approximating solutions to computationally expensive optimization problems. While using multiple machines to perform such computations in parallel is attractive as a means to reduce the running time, execution in a cluster environment requires substantial software infrastructure to cope with the challenges of a distributed system. In this paper, we introduce MRAnneal, a framework that simplifies the implementation of parallel simulated annealing algorithms. MRAnneal allows users to explicitly trade-off running time and the quality of approximate solutions by supplying only a small number of automatically tuned parameters. Our experimental results demonstrate that implementing applications using MRAnneal is straightforward and that such implementations yield approximate solutions quickly, even for applications without intuitive serial approximation heuristics.