Temperature and deadline aware runtime resource management with workload prediction for heterogeneous multi-core platforms

Abstract

Contemporary embedded platforms necessitate advanced resource management techniques to effectively utilize their diverse computational resources. Usually these platforms encounter fluctuations in workloads, making workload prediction have the potential to enhance resource management efficiency. In addition, in modern multi-core systems, there is a discernible tendency for processing cores to decrease in size relative to power consumption. This reduction in core size contributes to higher power density within the chips, leading to elevated chip temperatures. Therefore, addressing the temperature issue becomes critical. This paper introduces a prediction-based and temperature-aware resource management heuristic designed to meet task deadlines, while simultaneously considering energy minimization. When evaluated on real-life workload traces, the proposed method achieves a 5.5% increase in acceptance rate with one-step-ahead prediction and an 8.9% increase with four-steps-ahead prediction in a temperature-aware context, compared to scenarios without prediction.