Successful life forms in nature adapt to changes in the environment in ways that improve the efficiency of their activities or enhance their chances for survival. In the context of computer software, opportunities for adaptation arise from changes in the computational needs of an application as it executes, changes in availability of resources, failure of hardware, etc. However, most software today does not adapt to such changes, so its efficiency and survivability may be far from optimal. This project will develop general principles for building software systems that can adapt to such changes, and will demonstrate the validity of those principles by building prototype applications.
Progress in this area can only by made by focusing on a particular application domain that requires adaptive software, and assembling a team of applications researchers and computer scientists to tackle the problems in that domain at all levels. This project brings together a strong interdisciplinary team to focus on Computational Field Simulation (CFS). In particular, it will focus on computational fracture mechanics and reactive, multiphase fluid flows, both of which have an enormous number of opportunities for adaptive methods. These adaptations can be classified into three distinct categories. Application-level adaptivity. A variety of mathematical models may be available to describe the science of a problem, and it may be advantageous to switch adaptively between them to trade accuracy for time or other resources. Algorithm-level adaptivity. There are often multiple algorithms for solving a given model (e.g. direct or iterative methods for solving linear equations), and it may be advantageous to switch adaptively between them to manage resource availability or properties of the desired output. System-level adaptivity. The computational environment may change (e.g. more processors may become available or some communications links may fail), and the computation must adapt to these changes or risk crashing or taking significantly more time than necessary. This project argues that a general architecture designed to exploit these opportunities must have a set of interoperable components on a software bus, hardware/software sensors for monitoring stimuli, and control modules for orchestrating the components in response to the stimuli. It will undertake algorithmic research in each of the component areas, and will synthesize complete adaptive codes from the components using ideas similar to those in hardware synthesis. Adaptivity will be exploited at all levels in the resulting software. The project finally hopes to abstract a general theory of adaptive software systems from its experience in building such systems for field-driven simulations.
