Technology Library

Mar 2006
Empirical Mapping of Abstract Executable Specifications to Networked and Closely Coupled Processing Systems

Abstract. In automotive systems, (i) distributed, real-time algorithms, operating on a set of networked electronic control units (ECUs), are used to control the car (for example: the stability system); and (ii) fine grained video and audio algorithms are often mapped to an ensemble of processors in order to achieve the required performance at a known cost and, for untethered devices, with a known power budget (for example: an after-market cell phone/PDA supporting video and audio streaming – terrestrial and satellite, GPS, etc.).   The mapping of complex algorithms to systems containing multiple processors that satisfy some set of optimality conditions is notoriously difficult [8]. An empirical, refutation-based (scientific) process; design of experiments technology; and multi-variate statistics [6] can be employed to help drive the optimization process. Experimentation involves performing potentially many hundreds of experiments, each of which is meticulously measured, so that the processed data can be used to make decisions about the next steps to take in modifying the system (hardware, algorithms, software mappings, input/output interfaces, etc.) in the iterative march towards an optimal system. Undertaking such an experimental regime using physical hardware is prohibitively expensive in time and resources. The use of virtual (modeled) systems, together with the statistical machinery and empirical processes, makes this process plausible and necessary. This strategy is somewhat at variance with the AutoSAR & JASPAR standardization initiatives but is consistent with the notion of services-based interfaces where, for example, the always best external data sources (such as satellite and terrestrial RF data) – whether captured through the car’s infrastructure or a plugged-in external unit – should be available to whatever process is providing a requested service.