Thermal-aware joint CPU and memory scheduling for hard real-time tasks on multicore 3D platforms

Abstract

Designing 3D systems with on-chip DRAM is a promising solution to improve memory bandwidth and reduce memory access latency. However, 3D chips exacerbate the chip thermal problem due to their longer heat dissipation path, as well as the tight thermal coupling between logic and memory layers. In this paper, we are interested in studying thermal aware resource management strategies for both CPUs and memory systems when realizing hard real-time systems on 3D platforms under given peak temperature constraints. Given the dramatically increased power density not only from CPUs but also from memory systems as well, we believe that a joint CPU and memory system resource management is highly desired for 3D platforms to effectively deal with the heat dissipation confined in a small package. In addition, different from many existing thermal management strategies, which are reactive and best-effort in nature, we are more interested in ones that can ensure the strong guarantee for real-time applications. To this end, we introduce a novel approach that incorporates the periodic resource model to guarantee timing constraints for hard real-time systems under thermal constraints. In the meantime, by periodically (deterministically) throttling the accesses of CPUs and memory resources, our approach can effectively guarantee the thermal constraints imposed on both CPUs and memory systems. We use simulation results to demonstrate the effectiveness of our proposed approach in guaranteeing both the timing and temperature constraints for hard real-time tasks on 3D platforms.