A simulation methodology for reliability analysis in multi-core SoCs

Ayse K. Coskun, Tajana Simunic Rosing, Yusuf Leblebici, Giovanni De MicheliACM Great Lakes Symposium on VLSI 2006 · DOI

TL;DR

This paper introduces a novel simulation methodology for analyzing the reliability of multi-core Systems-on-Chip (SoCs), focusing on temperature-induced failures. By bridging the gap between high-level statistical models and detailed architectural simulations, it enables fine-grained, system-level reliability analysis, crucial for modern, power-constrained designs.

What Problem Does This Paper Address?

As semiconductor technology scales to deeper submicron and nanoscale levels, integration density skyrockets, leading to increased power densities and, consequently, higher operating temperatures. This elevated temperature is a primary driver of hard faults in electronic devices, significantly impacting system reliability. While techniques like Dynamic Power Management (DPM) and Dynamic Voltage Scaling (DVS) aim to reduce power consumption and thus temperature, aggressive application of these policies can paradoxically increase reliability issues due to temperature cycling. Existing reliability analysis tools are either too abstract (statistical models) or too computationally expensive (architecture-level models) to effectively analyze complex multi-core SoCs. This paper addresses the need for a simulation methodology that can accurately model and analyze reliability at a core level, considering both design choices and runtime events.

Methods

The authors propose a simulation methodology that combines several key components:

  • Workload Modeling: They generate synthetic workloads using a random task generation method, assuming independent periodic tasks. This allows for exploring various workload scenarios and their impact on system behavior.
  • Power Management: The simulator integrates DPM and DVS strategies. Specifically, it implements a fixed-timeout DPM policy and a DVS strategy derived from utilization updating techniques. These policies dynamically adjust power states and frequencies based on workload.
  • Thermal Modeling: The methodology leverages the HotSpot simulator, a well-established tool for architectural-level thermal modeling. HotSpot models thermal resistances and capacitances, accounting for lateral heat diffusion on the chip and detailed thermal package models.
  • Failure Rate Modeling: The core of the reliability analysis lies in modeling temperature-induced hard failures. The paper considers three primary mechanisms:
  • Electromigration (EM): Modeled using Black's model, which relates failure rate to current density and temperature.
  • Time-Dependent Dielectric Breakdown (TDDB): Modeled using a field-driven model, considering electric field and temperature.
  • Thermal Cycling (TC): Modeled using the Coffin-Mason equation, focusing on fast thermal cycles caused by frequent power state transitions. A key innovation here is the use of a sliding window to capture temperature and frequency data for TC computation, allowing for more accurate analysis of runtime temperature variations compared to lifetime-averaged approaches.
  • System Reliability Calculation: The system failure rate is computed using the Sum-of-Failure-Rates (SOFR) model, assuming independent failure rates per core.

The novelty lies in the integration of these components into a unified, event-driven simulation framework that samples temperature and failure rates at a granularity comparable to power management policy decisions, enabling fine-grained, long-term reliability analysis.

Datasets

The paper does not explicitly mention specific benchmark datasets. Instead, it relies on:

  • Synthetic Workloads: Generated using a random task generation method, with task periods and worst-case execution times (WCET) assigned based on typical values found in literature.
  • XScale Power Data: Power and frequency data for cores are based on Intel XScale processors, providing realistic parameters for simulation.
  • Thermal Package Data: Input parameters for the HotSpot simulator are taken from industry data (e.g., from [5], [13]) and specific values from [18].

Key Insights and Ideas

  • The Double-Edged Sword of Power Management: While DPM and DVS reduce average power and temperature, aggressive policies can exacerbate reliability issues through thermal cycling. This highlights the need for a balanced approach.
  • Fine-Grained Analysis is Crucial: Traditional architectural models are too slow, and statistical models too coarse. The proposed methodology's ability to sample at the same granularity as power management decisions is key to capturing dynamic reliability variations.
  • Thermal Cycling is a Major Concern: Fast thermal cycles, driven by frequent power state changes, significantly impact reliability and require careful modeling, especially with sliding window techniques.
  • Design Choices Matter: Thermal packaging and core placement are critical design factors that can significantly influence system reliability by affecting heat dissipation and hot spot distribution.
  • Workload Balancing is Key for Reliability: Distributing workloads to promote heat sharing among cores can mitigate the negative reliability impacts of power management.

Key Results

  • Power Management Trade-offs: For a 30-task set, DPM achieved 30% power savings but led to a decrease in Mean Time To Failure (MTTF). DVS with sleep offered over 40% savings and better reliability.
  • Workload Impact: In a batch workload scenario with DPM, average system MTTF decreased from 34 to 24 years. A producer-consumer workload scenario improved MTTF by 23% on average compared to batch, demonstrating the benefit of workload distribution for heat sharing.
  • Core Placement Benefits: Spacing out hot cores by introducing cooler structures between them increased the MTTF of a hot core by 8%, suggesting that thermal-aware placement is crucial.
  • Thermal Package Influence: Simple thermal packages (without heat sinks) resulted in significantly lower MTTF compared to more sophisticated ones, highlighting the importance of adequate cooling solutions for reliability.

Limitations and Future Work

  • Limitations:
    • The paper focuses on intrinsic hard failures and does not explicitly discuss soft errors.
  • The simulation assumes independent failure rates for cores in the SOFR model, which might not always hold true in complex systems.
  • The study uses synthetic workloads, and while representative, real-world application traces might yield different results.
  • Future Work:
    • The authors suggest that the simulation methodology can be extended to analyze other failure mechanisms and to explore more complex system architectures.
  • Further investigation into joint optimization of power management, thermal management, and workload scheduling for enhanced reliability is implied.
  • The methodology could be used to develop more sophisticated runtime reliability management strategies.

Summary updated

BibTeX
@inproceedings{coskun2006simulation,
  author = {Ayse K. Coskun and Tajana Simunic Rosing and Yusuf Leblebici and Giovanni De Micheli},
  title = {{A simulation methodology for reliability analysis in multi-core SoCs}},
  booktitle = {ACM Great Lakes Symposium on VLSI},
  year = {2006},
  doi = {10.1145/1127908.1127933},
}
Figure