Loading...

Design and Implementation of a Watchdog Processor using BMC for Remote Reliability Management

Ghasempour, Amin | 2025

21 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58822 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Ejlali, Alireza
  7. Abstract:
  8. Predictive Maintenance (PdM) plays a pivotal role in ensuring the continuous operation and reducing the operational costs of computer servers. The success of this strategy is directly contingent upon the integrity and accuracy of hardware sensor data. The emergence of faults within these sensors provides erroneous information, thereby undermining the integrity of PdM systems. This can lead to costly, unnecessary maintenance decisions or, conversely, the oversight of impending failures. This research introduces a novel, purely software-based solution for concurrent fault detection in server sensors. By intelligently leveraging the Baseboard Management Controller (BMC) as a watchdog processor, this approach obviates the need for any additional hardware or modifications to the system architecture. The proposed method employs a dedicated Kalman filter for each sensor to generate residuals (prediction error) and subsequently detects fault occurrences by statistically analyzing these residuals within a moving time window. A key innovation of this research is a comprehensive, systematic, and data-driven training process. Relying solely on fault-free operational data, this process uniquely estimates all critical system parameters for each sensor. These parameters include the Kalman filter model ($F, Q, R$), identified using Maximum Likelihood Estimation; the optimal analysis window size ($W^*$), determined through an evaluation of statistical stability; and the fault detection thresholds, which are derived from the statistical behavior of the residuals. The proposed method was implemented and integrated into the open-source OpenBMC firmware. The system's performance was rigorously evaluated within a comprehensive and realistic framework, utilizing QEMU-based emulation of the Palmetto server platform. By injecting constant, transient, and noise faults into real-world data from the processor's temperature sensor, the system's efficacy was assessed. The results demonstrate that the system achieved an overall Fault Detection Rate (FDR) of $94.46\%$, a False Alarm Rate (FAR) of $0.00\%$, and a Mean Time To Detection (MTTD) of $30.16$ seconds
  9. Keywords:
  10. Fault Detection ; Kalman Filters ; Watchdog Processors ; Server Processors ; Predictive Maintenance ; Baseboard Management Controller (BMC)System ; Data Driven Parameter Estimation

 Digital Object List

 Bookmark

...see more