Composite construction of high performance scientific applications
Author(s)
Mayer, Anthony E.
Type
Thesis
Abstract
Scientific computing applications require both high performance and structured software engineering to allow their development by non-computing specialists. Traditional component-based design patterns provide a separation between interface and implementation that supports software engineering, but neither enables high performance nor exploits the scientist's knowledge of the application domain. In this thesis we present a refinement of the component based design paradigm, where the binding between abstraction and implementation is deferred until run-time. By extending the abstraction to include meta-data regarding the component's. behaviour and dependencies it is possible to perform run-time optimisation and verification. Scientific knowledge such as the numerical stability constraints on a simulation may be used to verify the semantic correctness of the composition during application construction, and may be retained until run-time in order to parameterise the application. With the binding between implementation and abstraction being deferred until run-time, an abstraction may possess many different implementations. Such multiple implementations enable the deployment of the application upon heterogenous resources, by selecting the correct implementation for the context. Such selection is guided by the estimated performance of the composite application. This composite performance modelling is only possible through the use of the high-level behavioural component meta-data. Hence the meta-data supports not only accessibility of design, but enables high performance. Complex scientific examples are given which demonstrate how such benefits may be obtained within a realistic distributed environment. A structured language for recording the meta-data and component design has been developed, together with a prototype framework that utilises this information to inform run-time optimisation and implementation selection.
Version
Open Access
Date Issued
2002
Date Acceptance
2002
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)