High level thermal-aware scheduling for multiprocessors

• Main Author: Jin, Cui
• Other Authors: Douglas Leslie Maskell
• Format: Theses and Dissertations
• Language: English
• Published: 2013
• Subjects: DRNTU::Engineering::Computer science and engineering::Software::Operating systems; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Simulation and modeling; DRNTU::Engineering::Computer science and engineering::Computer systems organization::Performance of systems; DRNTU::Engineering::Computer science and engineering::Computer systems organization::Processor architectures
• Online Access: https://hdl.handle.net/10356/54996
• Institution: Nanyang Technological University

Power and thermal issues are primary design constraints in both stationary and portable computing devices. Adverse thermal issues can impact microprocessor performance, including computational speed degradation, aging, and unreliable system behaviour. These situations are exaggerated in current state-of-the-art multiprocessors due to their high power density and the thermal coupling between cores. High level thermal-aware scheduling (TAS) is seen as one possible solution to optimize and control on-chip temperature. However, after performing an extensive review of the literature, a number of shortcomings in current high level TAS implementations have been identified. These include, the inaccuracy of thermal sensor readings, low computational efficiency of existing time-triggered thermal simulators, oversimplified thermal and leakage power models currently used at the system level, lack of appropriate thermal constraints used in scheduling analysis in hard real-time embedded systems and a lack of appropriate fine-grained dynamic TAS (DTAS). These shortcomings have provided suitable motivation for the work described in this thesis, which includes the following contributions: • A fast event-driven look-up table (LUT) based thermal estimation approach is developed. We introduce the concept of power events which capture the significant power changes on-chip. These power events induce a temperature change which can be easily obtained using the pre-calculated LUTs (representing the thermal response of a unit power input). We show that these thermal responses, induced by individual power events, satisfy the superposition principle and can be accumulated to evaluate the thermal map when any event occurs. We also define the necessary optimizations and operations for the LUTs. Experimental results show our LUT method is accurate, producing thermal estimations of similar quality to an existing open-source thermal simulator (HotSpot), while providing 2 to 3 orders of magnitude reduction in computational complexity. • We use our fast LUT approach to analyze the offline schedulability for a real-time task set on a simulated multiprocessor system under a strict (hard) thermal constraint. This is very useful for reducing the risk of overheating in safety-critical embedded systems. Our schedulability test can also be used as a framework to optimize other goals (e.g. maximizing the performance and minimizing the peak temperature). We show that we are able to schedule large task sets (up to 50 tasks) in reasonable time (less than 12 minutes), which is 2 to 3 orders of magnitude faster than using scheduling with existing thermal simulation tools. • For high power multiprocessor (or many-core) systems, it is not possible to ignore the temperature-leakage power dependence. Therefore, we modify the LUT-based approach to include a temperature-dependent leakage power model. The leakage power calibration enables us to accurately predict the near future thermal map without needing to resort to a computationally expensive iterative approach. Based on this prediction, we develop several heuristic policies for dynamic TAS on a simulated many-core system. We show that our proposed predictive policies are significantly better, in terms of minimizing average/peak temperature, reducing the dynamic thermal management overhead and improving other real-time features, than existing TAS schedulers, making them highly suitable for heuristically guiding thermal aware task allocation and scheduling.