Augmented Cross-Entropy-Based Joint Temperature Optimization of Real-Time 3-D MPSoC Systems
Yangguang Cui,Kun Cao,Liying Li,Junlong Zhou,Tongquan Wei,Shiyan Hu
DOI: https://doi.org/10.1109/tcad.2019.2939328
IF: 2.9
2019-01-01
IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Abstract:3-D multiprocessor system-on-chip (MPSoC) systems can offer higher integration density, lower interaction cost, better bandwidth, and greater performance. However, vertically stacked silicon layers and limited heat dissipation paths result in high peak temperature and large temperature variation, which incur reliability reduction, lifetime decay, and performance degradation. In this article, we propose an offline augmented cross-entropy (CE)-based task scheduling strategy to jointly optimize peak temperature and temperature variation under the constraint of timeliness. Specifically, based on the conventional CE method, a heuristic iterative sampling method is designed to explore task-to-core assignment for balanced heat distribution between the top-layer and the bottom-layer cores. Subsequently, thermal characteristics of 3-D MPSoC systems are used to judiciously swap tasks between the two layers to improve the conventional CE-based task assignment and accelerate the iterative process. The peak temperature of individual cores is further reduced via sequencing, splitting, and slacking task execution. The experimental results demonstrate that compared to the existing state-of-the-art methods, the proposed scheme can reduce peak temperature by up to 8.02 °C and temperature variation by up to 24.78% without violating the timeliness of tasks.