A Machine Learning Based Approach for Fast Demand Characterization of Digraph Tasks

Key Speaker

Abstract

Many complex real-time applications such as self-driving cars can be modeled using digraph tasks. A digraph task consists of vertices, which denote computations, and directed edges, which capture inter-task dependencies. In contrast to simpler real-time task models such as the sporadic task model, digraphs are more expressive and can model code structures such as loops. However, calculating the worst-case demand of a digraph can be time consuming, especially during online demand recharacterization. Inspired by recent advances in and successes of ML algorithms in solving problems in other application domains, we present an ML-based demand characterization framework for digraph tasks. Our framework is experimentally shown to provide accurate, safe, and fast demand (re)characterization compared to state-of-the-art techniques, which is crucial as real-time systems and cyber-physical systems increasingly become more complex and may evolve during runtime.

Authors

1. Rajarshi Mukherjee rajarshim13@vt.edu (Virginia Polytechnic and State University)
2. Thidapat (Tam) Chantem tchantem@vt.edu (Virginia Polytechnic Institute and State University)

 Overview  Program