Application of Dempster-Shafer Theory to task mapping under epistemic uncertainty

A common problem in system design is a lack of knowledge about the system parameters in early design stages. This results in epistemic uncertainty in the systems performance. Classic probability theory imposes unfavourable restrictions for the evaluation of different system realisations under epistemic uncertainty. An alternative mathematical approach is Dempster-Shafer Theory. In this paper we investigate the integration of Dempster-Shafer Theory in the Longest Processing Time algorithm, which is a heuristic for task mapping in embedded system design. The algorithm accepts uncertain estimates of processor speeds and task complexities. It can produce several plausible task mappings based on a degree of pessimism. We propose two criteria to compare these mappings in terms of performance and risk. Moreover we propose an approximation for arithmetic operations on Dempster-Shafer structures, which makes these tractable.