Thermal aware overall energy minimization scheduling for hard real-time systems

Huang Huang, Ming Fan, Gang QuanSustain. Comput. Informatics Syst 2013 · DOI

TL;DR

This paper tackles the challenge of minimizing energy consumption in hard real-time systems by considering both dynamic and leakage power, crucially accounting for their interdependence with temperature. They introduce a novel, efficient energy estimation method and two scheduling algorithms (off-line and on-line) that significantly outperform existing approaches in reducing energy consumption while guaranteeing real-time deadlines.

What Problem Does This Paper Address?

Modern computing systems, especially those in the deep sub-micron era, face a dual challenge: increasing power density and the growing significance of leakage power. As transistors shrink, leakage current becomes a substantial contributor to overall energy consumption, and this leakage is highly sensitive to temperature. This temperature dependency creates a complex feedback loop: higher temperatures increase leakage, which in turn generates more heat. For battery-powered embedded systems, this is a critical issue as battery technology lags behind semiconductor advancements. For server systems, it leads to increased cooling costs and reduced reliability.

Existing energy minimization techniques often focus solely on dynamic power or make simplifying assumptions about leakage and temperature. This paper addresses the need for a more holistic approach that accurately models and minimizes overall energy consumption (dynamic + leakage) in hard real-time systems, where meeting deadlines is paramount. The core problem is how to schedule tasks to minimize energy while ensuring all deadlines are met, considering the intricate relationship between voltage, temperature, and leakage.

Methods

The authors propose a two-pronged approach:

  1. Novel Energy Estimation Method: They develop a closed-form energy estimation equation that accurately and efficiently calculates the total energy consumption of a given speed schedule. This method directly accounts for the leakage/temperature dependency. It's a significant improvement over existing methods that rely on dividing intervals into many small sub-intervals, which is computationally expensive. The method leverages a thermal model and a leakage power model that considers both voltage and temperature.
  2. Energy Minimization Scheduling Algorithms: Based on the new energy estimation method, they introduce two scheduling algorithms:
  3. Off-line Algorithm (for single periodic task): This algorithm targets systems with a single periodic task (or multiple tasks with the same period and deadline) and aims to minimize energy consumption under thermal steady-state conditions. It builds upon the "M-Oscillating" technique, which involves switching between two neighboring speeds and idle modes to manage temperature. The authors extend this by incorporating transition overheads and using their energy estimation method to find the optimal oscillation parameters (m, the number of divisions) for minimal energy.
  4. On-line Algorithm (for multiple sporadic tasks): This algorithm is designed for more general real-time systems with multiple sporadic tasks scheduled using the Earliest Deadline First (EDF) policy. It also utilizes the M-Oscillating principle but adapts it for on-line decision-making. At each scheduling point, it calculates the required constant speed to meet the deadline, identifies neighboring speeds, and then uses their energy estimation method to find the optimal oscillation parameters (m) that minimize energy consumption for the current job.

Datasets

Not discussed in the paper. The authors use simulated processor models and parameters derived from existing literature and technology specifications (e.g., 65 nm technology).

Key Insights and Ideas

  • Leakage-Temperature Interdependence is Crucial: The paper strongly emphasizes that ignoring the coupled relationship between leakage power and temperature leads to significant inaccuracies and ineffective energy optimization in modern systems.
  • Closed-Form Energy Estimation is Key: Developing an accurate and computationally efficient energy estimation method is fundamental. Their closed-form solution dramatically reduces the overhead compared to interval-splitting approaches.
  • M-Oscillating for Thermal Management: The M-Oscillating technique, when combined with accurate energy estimation, proves effective for both off-line and on-line energy minimization by actively managing temperature.
  • On-line Scheduling for Dynamic Environments: For complex, dynamic systems with sporadic tasks, an on-line approach is essential because thermal steady-state assumptions are often invalid. Their OLMO algorithm provides a practical solution.
  • Trade-offs in Scheduling: The work highlights the trade-off between energy savings and computational overhead, especially for on-line algorithms. Their approach aims to find an optimal balance.

Key Results

  • Energy Estimation Speedup: The proposed energy estimation method achieves speedups of up to two orders of magnitude (e.g., 177x) compared to existing interval-splitting methods, while maintaining high accuracy (maximum relative error < 2.7% for single intervals, < 4.1% for transient states).
  • Off-line Scheduling Performance: The proposed off-line algorithm significantly outperforms naive and pattern-based approaches in energy reduction (e.g., 26.5% vs. 13.2% over naive) and peak temperature reduction (e.g., 21.8°C lower than naive).
  • On-line Scheduling Performance: The OLMO algorithm demonstrates substantial energy savings compared to pattern-based and OLDVS approaches, saving over 10% on average and up to 20% at high system utilization. These savings are attributed to better temperature management and reduced leakage energy.

Limitations and Future Work

  • Model Simplifications: The paper acknowledges that the leakage model uses linear approximation for temperature dependency, which, while maintaining accuracy, is still an approximation. The thermal model is also a simplified RC model.
  • Single-Core Focus: The current work focuses on single-core systems. Extending the algorithms to multi-core architectures, which introduce more complex thermal interactions, is a natural next step.
  • Task Model Assumptions: The off-line algorithm assumes periodic tasks with known WCET and steady-state conditions. The on-line algorithm handles sporadic tasks but still relies on WCET for initial calculations. Future work could explore more dynamic task execution times.
  • Transition Overhead: While accounted for, the exact nature and variability of transition overheads could be further investigated.

The authors suggest that future work could explore extending these techniques to multi-core systems and investigating more sophisticated searching algorithms for finding optimal parameters in on-line scenarios.

Summary updated

BibTeX
@article{huang2013thermal,
  author = {Huang Huang and Ming Fan and Gang Quan},
  title = {{Thermal aware overall energy minimization scheduling for hard real-time systems}},
  journal = {Sustain. Comput. Informatics Syst.},
  year = {2013},
  doi = {10.1016/j.suscom.2013.06.002},
}
Figure