Computational Quality of Service in Parallel CFD

Component-based environments provide opportunities to improve the performance, numerical accuracy, and other characteristics of parallel simulations in computational fluid dynamics (CFD). Because component-based software engineering combines object-oriented design with the powerful features of well-defined interfaces, programming language interoperability, and dynamic composability, it helps to overcome obstacles that make it difficult to share even welldesigned traditional numerical libraries. Not only can applications be assembled from components selected to provide good algorithmic performance and scalability, but they can also be changed dynamically during execution to optimize desirable characteristics. This concept of the automatic selection and configuration of components to suit a particular computational purpose is called computational quality of service (CQoS) [7, 13]. CQoS embodies the familiar concept of quality of service in networking and the ability to specify and manage characteristics of the application in a way that adapts to the changing (computational) environment. The factors that affect performance are closely tied to a component’s parallel implementation, its management of memory, the algorithms executed, the algorithmic parameters employed (for example, the level of overlap in an additive Schwarz preconditioner), and other operational characteristics. Scientific component software is also concerned with functional qualities, such as the level of accuracy achieved for a particular algorithm. This paper presents an overview of new infrastructure for automated performance gathering and analysis of high-performance components, which is a key facet of our CQoS research, with emphasis on using these capabilities in parallel CFD simulations, such as flow in a driven cavity and compressible Euler flow. The remainder of this paper is organized as follows. Section 2 discusses parallel CFD applications and algorithms that motivate this infrastructure. Section 3 introduces the new framework for enabling CQoS in parallel nonlinear PDE-based applications, while Section 4 discusses conclusions and opportunities for future work.