Supporting adaptive and irregular parallelism for non-linear numerical optimization

- A global optimization framework for SMPs and multicore clusters is presented.
- It exploits hierarchal and dynamic task parallelism of the Multistart method.
- Gradient/Hessian calculations and Newton's optimization method are parallelized.
- Several task distribution schemes are studied and evaluated.
- Our framework is applied successfully to the protein conformation problem.

This paper presents an infrastructure for high performance numerical optimization on clusters of multicore systems. Building on a runtime system which implements a programming and execution environment for irregular and adaptive task-based parallelism, we extract and exploit the parallelism of a Multistart optimization strategy at multiple levels, which include second order derivative calculations for Newton-based local optimization. The runtime system can support a dynamically changing hierarchical execution graph, without any assumptions on the levels of parallelization. This enables the optimization practitioners to implement, transparently, even more complicated schemes. We discuss parallelization details and task distribution schemes for managing nested and dynamic parallelism. In addition, we apply our framework to a real-world application case that concerns the protein conformation problem. Finally, we report performance results for all the components of our system on a multicore cluster.