On-Line Temperature Estimation for Noisy Thermal Sensors Using a Smoothing Filter-Based Kalman Predictor

Xin Li, Xingtao Ou, Zhi Li, Henglu Wei, Wei Zhou, Zhemin Duan

TL;DR

This paper tackles the critical issue of inaccurate temperature readings from on-chip sensors in multi-core processors, which hinders effective dynamic thermal management (DTM). They propose a novel "Multi-Sensor Synergistic Calibration Algorithm" (MSSCA) that combines a smoothing filter-based Kalman predictor with spatial correlation analysis to provide more accurate real-time temperature estimations, leading to better DTM decisions.

What Problem Does This Paper Address?

Modern multi-core processors generate significant heat, and managing this heat is crucial for performance, reliability, and energy efficiency. Dynamic Thermal Management (DTM) systems rely on on-chip thermal sensors to monitor temperatures and adjust processor behavior (like voltage and frequency) to prevent overheating. However, these embedded sensors are notoriously noisy due to factors like manufacturing variations, voltage fluctuations, and environmental changes. This noise leads to discrepancies between the sensor readings and the actual chip temperatures, causing DTM systems to make suboptimal or incorrect decisions, such as unnecessary throttling or failing to react to critical temperature spikes. The paper aims to bridge this gap by providing a robust method for accurately estimating true chip temperatures from these noisy sensor inputs.

Methods

The core of the proposed solution is a two-pronged approach:

  1. Smoothing Filter-Based Kalman Predictor: This is a standard technique for state estimation in systems with noisy measurements. The paper adapts it by:
  2. Polynomial Fitting: First, they establish a non-linear relationship between the output frequency of ring oscillators (a common type of on-chip sensor) and temperature using polynomial fitting. This helps to linearize the sensor's response.
  3. Smoothing Filter: A simple moving average (SMA) is used as a smoothing filter to reduce the impact of short-term temperature fluctuations and improve the prediction model.
  4. Kalman Filter: This filter then combines the smoothed predictions with the noisy sensor measurements to produce a more accurate, real-time temperature estimate for individual sensors.
  5. Multi-Sensor Synergistic Calibration Algorithm (MSSCA): This is the novel contribution, designed to leverage information from multiple sensors.
  6. Spatial Correlation Model: The algorithm exploits the fact that temperatures at nearby sensor locations are correlated. They use a mathematical model (based on Bessel functions) to quantify this spatial correlation between sensors based on their physical distance on the chip.
  7. Synergistic Calibration: MSSCA uses these correlation coefficients to "correct" the noisy readings of one sensor by considering the readings and estimates of its correlated neighbors. If a sensor's reading deviates significantly from its neighbors, it's adjusted. This process is iterated across all sensors to improve their simultaneous estimation accuracy.

Datasets

The paper does not explicitly mention training datasets. Instead, they introduce a novel experimental setup for data acquisition:

  • Infrared Imaging-Based Temperature Measurement: They developed a system using an oil-based cooling system and an infrared-transparent heat sink to capture real-time thermal traces of an AMD quad-core processor. This provides ground truth temperature data for evaluation.
  • AMD Quad-Core Processor Benchmarks: They tested their approach on an AMD Athlon II X4 610e processor running various SPEC CPU 2006 benchmark workloads. This allows for evaluation under realistic operating conditions.

Key Insights and Ideas

  • Noise is a Major Hurdle: The paper strongly emphasizes that ignoring sensor noise leads to significant errors in DTM, potentially causing missed thermal events or false alarms.
  • Spatial Correlation is Key for Multi-Sensor Systems: The core innovation lies in recognizing and exploiting the spatial relationships between sensors. Nearby sensors provide valuable context for correcting each other's noisy readings.
  • Kalman Filters are Powerful, but Need Enhancement: While Kalman filters are effective for single-sensor denoising, they don't inherently account for inter-sensor dependencies. MSSCA builds upon this foundation.
  • Real-World Validation is Crucial: The use of infrared thermography to capture actual chip temperatures provides a more realistic evaluation than purely simulation-based approaches.

Key Results

The proposed MSSCA algorithm demonstrated significant improvements over raw sensor readings and a standard Kalman filter:

  • Reduced RMSE: The MSSCA achieved a reduction of up to 1.2 °C in Root-Mean-Square Error (RMSE) compared to ideal sensor readings (when noise standard deviation was 10%).
  • Increased SNR: The Signal-to-Noise Ratio (SNR) was improved by up to 15.8 dB compared to ideal sensor readings.
  • Lower False Alarm Rate (FAR): The average FAR was reduced by 28.6%, meaning DTM systems are less likely to trigger unnecessarily or miss critical events.
  • Runtime Efficiency: The MSSCA has a small average runtime overhead, making it suitable for on-line estimation within the processor's sampling interval.

Limitations and Future Work

  • Correlation Model Parameters: The paper mentions that the spatial correlation function's shape is regulated by parameters 'b' and 's', which were set to 1 and 8, respectively. The sensitivity of the algorithm to these parameters and how they are optimally determined for different chip architectures could be a point for future exploration.
  • Memory Requirement: The MSSCA requires additional memory to store the correlation coefficients among all sensors. While deemed acceptable, this could be a consideration for highly resource-constrained environments.
  • Generalizability: While tested on an AMD processor, further validation on different architectures and process nodes would be beneficial to confirm the broad applicability of the MSSCA.
  • Dynamic Correlation: The current model assumes correlation is primarily distance-dependent and fixed. Future work could explore how correlations might change dynamically with varying workload distributions or thermal gradients.

Summary updated

Abstract

Dynamic thermal management (DTM) mechanisms utilize embedded thermal sensors to collect fine-grained temperature information for monitoring the real-time thermal behavior of multi-core processors. However, embedded thermal sensors are very susceptible to a variety of sources of noise, including environmental uncertainty and process variation. This causes the discrepancies between actual temperatures and those observed by on-chip thermal sensors, which seriously affect the efficiency of DTM. In this paper, a smoothing filter-based Kalman prediction technique is proposed to accurately estimate the temperatures from noisy sensor readings. For the multi-sensor estimation scenario, the spatial correlations among different sensor locations are exploited. On this basis, a multi-sensor synergistic calibration algorithm (known as MSSCA) is proposed to improve the simultaneous prediction accuracy of multiple sensors. Moreover, an infrared imaging-based temperature measurement technique is also proposed to capture the thermal traces of an advanced micro devices (AMD) quad-core processor in real time. The acquired real temperature data are used to evaluate our prediction performance. Simulation shows that the proposed synergistic calibration scheme can reduce the root-mean-square error (RMSE) by 1.2 ∘ C and increase the signal-to-noise ratio (SNR) by 15.8 dB (with a very small average runtime overhead) compared with assuming the thermal sensor readings to be ideal. Additionally, the average false alarm rate (FAR) of the corrected sensor temperature readings can be reduced by 28.6%. These results clearly demonstrate that if our approach is used to perform temperature estimation, the response mechanisms of DTM can be triggered to adjust the voltages, frequencies, and cooling fan speeds at more appropriate times.
BibTeX
@inproceedings{li2018line,
  author = {Xin Li and Xingtao Ou and Zhi Li and Henglu Wei and Wei Zhou and Zhemin Duan},
  title = {{On-Line Temperature Estimation for Noisy Thermal Sensors Using a Smoothing Filter-Based Kalman Predictor}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=g5SrB009oj},
}
Figure