IMPACT: Integrated Multilevel Performance Framework for Scalable Systems

The current state-of-art in performance technology does not facilitate detailed performance evaluation of complex, large scale systems. The aim of this proposal is to design IMPACT, an integrated performance framework that can be used to measure the performance of operational systems and to extrapolate those measurements into performance predictions for alternative machine configurations. IMPACT will use multi-level, hierarchical models to support end-end performance models of systems that include application level, system software level, and hardware level models. Where feasible, direct execution will be used to integrate measurements on operational system components. The framework will support modeling of subsystems at multiple levels of granularity, and the composition of a model from such multi-level models. It will also incorporate parallel simulation technology as a modular option such that parallelism can be explored in an iterative manner during the model development cycle. The Phase I report will produce a detailed design and requirements analysis for the framework. To demonstrate its feasibility, a simple library of models will be prototyped to simulate hardware, software, and application levels of a parallel database system that can simulate sample queries from the TPC-D benchmark suite.