RePAiR: A Strategy for Reducing Peak Temperature while Maximising Accuracy of Approximate Real-Time Computing: Work-in-Progress

Abstract

Improving accuracy in approximate real-time computing without violating thermal-energy constraints of the underlying hardware is a challenging problem. The execution of approximate real-time tasks can individually be bifurcated into two components: (i) execution of the mandatory part of the task to obtain a result of acceptable quality, followed by (ii) partial/complete execution of the optional part, which refines the initially obtained result, to increase the accuracy without violating the temporal-deadline. This paper introduces RePAiR, a novel task-allocation strategy for approximate real-time applications, combined with fine-grained DVFS and on-line task migration of the cores and power-gating of the last level cache, to reduce chip-temperature while respecting both deadline and thermal constraints. Furthermore, gained thermal benefits can be traded against system-level accuracy by extending the execution-time of the optional part.