As computational science progresses toward ever more realistic multiphysics and multiscale applications, no single research group can effectively select or tune all components of a given application, and no solution strategy can seamlessly span the entire spectrum efficiently. Common component interfaces, along with programming language interoperability and dynamic composability, are key features of component technology that enable easy access to suites of independently developed algorithms and implementations; via the Common Component Architecture (CCA), such capabilities are now making inroads in scientific computing. The challenge then becomes how, during runtime, to make the best choices for reliability, accuracy, and performance.
Researchers in the CCA Working Group on Computational Quality of Service (CQoS) are addressing this challenge by developing tools for CQoS, or the automatic selection and configuration of components to suit a particular computational purpose and environment. The two main facets of CQoS tools are (1) measurement and analysis infrastructure and (2) control infrastructure for dynamic component replacement and domain-specific decision making.
Collaborations with four SciDAC applications teams motivate this work: parallel mesh partitioning in combustion (CFRFS project, BES), managing resources in quantum chemistry (BES), and efficiently solving linear systems in fusion and accelerator models (FACETS project, FES, and COMPASS project, HEP/NP/BES). For example, the quantum chemists recently developed a prototype component that automates the selection from among several integral evaluators in molecular computations; this approach achieved substantial performance improvement (greater than 25% in some preliminary tests) with respect to the worst-case configuration that a novice user might employ.
We are also collaborating with researchers in the Performance Engineering Research Institute (PERI) on the design and implementation of CQoS component infrastructure, in particular in the definition of common database formats and interfaces for representing performance data and associated metadata, both of which are required for implementing CQoS adaptive strategies.