Predictive Power Management for Multi-core Processors

William Lloyd Bircher, Lizy K. JohnISCA Workshops 2010 · DOI

TL;DR

This paper introduces a novel predictive power management technique for multi-core processors that leverages core activity prediction to anticipate workload phase changes. By accurately predicting these shifts, the system can proactively adjust power states, leading to significant performance improvements and reduced power consumption compared to reactive methods.

What Problem Does This Paper Address?

Modern multi-core processors are designed to dynamically adjust their performance capacity (voltage and frequency) to save power and improve efficiency. However, traditional "reactive" approaches, which adapt only after a workload's demand changes, suffer from a lag. This lag results in suboptimal performance during transitions and wasted power when demand decreases. This is particularly problematic for workloads with frequent and short phase changes, where reactive systems struggle to keep up. The paper aims to bridge this gap by developing a predictive system that can anticipate these changes before they occur, enabling immediate and optimal power/performance adaptations.

Methods

The core of the paper's contribution lies in its predictive approach. Instead of predicting metrics like Instructions Per Cycle (IPC) or memory-boundedness, which can lead to performance loss even when accurate, the authors focus on predicting core activity. Core activity is defined as the ratio of non-halted (active) cycles to total cycles. This is a crucial distinction because halted cores have zero performance dependence on frequency, meaning power can be saved without sacrificing performance.

They developed a table-based predictor that uses a Global Phase History Register (GPHR) and a Pattern History Table (PHT) to store and predict phase transitions based on past core activity. This predictor is designed to be lightweight and efficient.

To evaluate their approach, they also developed a detailed, run-time power model for an AMD quad-core processor. This model is notable for accounting for fine-grain temperature and voltage variations, as well as leakage power, which are often overlooked in simpler models. The model's coefficients are tuned empirically using real system measurements.

Datasets

The primary dataset used for evaluation is SYSmark® 2007. The authors highlight that SYSmark® 2007 is a more suitable benchmark for this study than traditional scientific/computing benchmarks (like SPEC CPU) because it better represents real-world desktop workloads. SYSmark® 2007 includes a mix of GUI interactions, I/O operations, thread migration, and a wide range of processor activity levels, leading to more complex and frequent phase changes. They also used SPEC CPU 2006 for validating their power model.

Key Insights and Ideas

  • Core Activity as the Key Predictor: The realization that predicting core activity (active vs. idle) is more effective for power management than predicting performance metrics like IPC is a significant insight. This is because idle cores offer a "free" opportunity for power savings without performance penalties.
  • SYSmark® 2007's Relevance: The paper effectively demonstrates why benchmarks like SYSmark® 2007, with their inherent complexity and frequent phase changes, are crucial for evaluating advanced power management techniques. Traditional benchmarks often mask these dynamic behaviors.
  • The Trade-off Between Prediction Accuracy and Coverage: The authors acknowledge that not all phases are predictable. Their approach intelligently combines high-accuracy prediction for predictable phases with reactive management for unpredictable ones, maximizing overall efficiency.
  • The Importance of a Detailed Power Model: Developing an accurate, fine-grained power model that accounts for temperature, voltage, and leakage is essential for realistically estimating the benefits of predictive power management.

Key Results

  • Prediction Accuracy: The predictor achieved an average 98% accuracy for 43% of the SYSMark2007 workload. The remaining portion was identified as unpredictable and handled reactively.
  • Performance Speedup: The proposed predictive approach, by enabling proactive power borrowing from idle cores, achieved an average speedup of 7.3% in SYSMark2007. The best-case speedup was 11.1% for the e-learning subtest.
  • Power Model Accuracy: The developed run-time power model demonstrated an average error rate of 0.89% across SPEC CPU 2006 and SYSMark2007.
  • Comparison to Reactive Schemes: The predictor outperformed reactive schemes in terms of core power accuracy, achieving 86% accuracy compared to the reactive scheme's 83%.

Limitations and Future Work

  • Predictability Limits: The paper acknowledges that a significant portion of the workload (57%) remains unpredictable by their current predictor.
  • Cross-Metric Prediction Error: The authors noted that predicting power indirectly using activity levels led to some aliasing and increased error compared to predicting activity directly.
  • Future Work: The paper suggests exploring cross-metric prediction (e.g., predicting power directly from activity) and investigating more sophisticated predictor structures to improve coverage of unpredictable phases. They also mention the need to re-tune the power model for processors with different semiconductor parameters.

Summary updated

BibTeX
@inproceedings{bircher2010predictive,
  author = {William Lloyd Bircher and Lizy K. John},
  title = {{Predictive Power Management for Multi-core Processors}},
  booktitle = {ISCA Workshops},
  year = {2010},
  doi = {10.1007/978-3-642-24322-6\_21},
}
Figure