Thermal Management and Communication Virtualization for Reliability Optimization in MPSoCs

Victor M. van Santen, Hussam Amrouch, Thomas Wild, Jörg Henkel, Andreas HerkersdorfEmbedded 2020 · DOI · OpenAlex

Abstract

This chapter covers the interdependencies of temperature and the reliable operation of multiprocessor systems on chip (MPSoC). Starting with the assessment of temperature values for the different cores either through measurement or estimation, it is shown which methods on system level can be applied to balance the thermal stress in the system and thus to come to an evenly distributed probability of errors. Special focus is given to task migration as a system-level means, which is supported by an on-chip interconnect virtualization technique that eases fast and transparent switch-over of communication channels. Overall, it is shown how mechanisms on different levels of the system stack can be combined to cooperate across layers for improving system reliability.

Introduction: The Challenge of Thermal Reliability

As semiconductor technology scales into the deep-submicron regime, Multi-Processor Systems on Chip (MPSoCs) have become increasingly vulnerable to thermal-induced failures. Heat acts as a "silent killer" through two primary mechanisms. First, it has a direct impact on circuit timing; higher temperatures slow down electron mobility, leading to immediate timing violations and data corruption. Second, heat accelerates physical aging phenomena, such as Bias Temperature Instability (BTI), which permanently degrades transistor performance over time.

The Impact of Temperature on Circuit Timing

Immediate reliability is often compromised when thermal hotspots emerge. For instance, in memory structures like SRAM, increased temperatures can shift the distribution of read access times, pushing them beyond the clock cycle limit and causing read failures.

Impact of temperature on SRAM read access time distribution.

As illustrated in the Monte Carlo simulations above, a temperature increase from 80°C to 125°C significantly shifts the read access time distribution, directly increasing the probability of failure ($P_{fail}$).

Methodology: A Cross-Layer Approach

To combat these challenges, the researchers propose a comprehensive cross-layer framework that bridges the gap between low-level physics and high-level system management. This approach involves:

  • High-Fidelity Thermal Sensing: Utilizing infrared (IR) thermography to achieve 50µm spatial resolution, overcoming the limitations of coarse-grained on-chip diodes.
  • Reliability Modeling: Linking physical transistor models (e.g., BSIM) to circuit-level simulations to quantify aging effects.
  • NoC Virtualization: Introducing Virtualized NoC Adapters (VNA) to decouple logical task addresses from physical core locations, enabling transparent task migration.

Modeling Long-Term Degradation

Beyond immediate timing issues, the framework accounts for long-term wear-and-tear. Bias Temperature Instability (BTI) causes a shift in the threshold voltage ($\Delta V_{Th}$) of transistors, a process that is highly sensitive to both time and temperature.

Aging-induced degradation (BTI) measured at various temperatures.

By modeling these physical shifts, the system can predict when a core is likely to fail and proactively migrate critical workloads to "healthier" or cooler parts of the chip.

System-Level Optimization: Migration and Redundancy

The framework employs advanced migration protocols, such as "Dualcast" and hardware-offloaded forwarding, to move tasks with near-zero downtime. This is complemented by Adaptive Modular Redundancy (AMR), which dynamically triggers Dual or Triple Modular Redundancy (DMR/TMR) for critical tasks when a core's health is predicted to decline.

One key strategy is "Non-Critical Task (NCT) ejection." When a core reaches a critical thermal threshold, the system evicts low-priority work to prioritize high-criticality tasks and prevent permanent hardware damage.

Key Results and Impact

The research quantifies a significant "aging tax": without active management, the probability of failure can increase by 42% over a system's lifetime. The proposed framework successfully mitigates this by:

  • Thermal Balancing: Evening out wear-and-tear across the MPSoC by migrating tasks away from hotspots.
  • Error Resilience: In stress tests involving 10,000 bit-flips, the protected system masked nearly all errors for high-criticality tasks, whereas unprotected systems suffered thousands of failures.
  • Performance Transparency: Hardware-offloaded migration protocols reduced processing latency, making task moves nearly unnoticeable to the application.

Conclusion

Optimizing reliability in modern MPSoCs requires a holistic view of the system stack. By combining high-resolution thermal sensing, physical aging models, and architectural virtualization, this framework provides a robust path toward self-healing, reliability-aware computing. Future work will focus on reducing the hardware overhead of virtualization and improving predictive models for aging-related failures.

Summary updated

Abstract

Abstract This chapter covers the interdependencies of temperature and the reliable operation of multiprocessor systems on chip (MPSoC). Starting with the assessment of temperature values for the different cores either through measurement or estimation, it is shown which methods on system level can be applied to balance the thermal stress in the system and thus to come to an evenly distributed probability of errors. Special focus is given to task migration as a system-level means, which is supported by an on-chip interconnect virtualization technique that eases fast and transparent switch-over of communication channels. Overall, it is shown how mechanisms on different levels of the system stack can be combined to cooperate across layers for improving system reliability.
BibTeX
@inproceedings{santen2020thermal,
  author = {Victor M. van Santen and Hussam Amrouch and Thomas Wild and Jörg Henkel and Andreas Herkersdorf},
  title = {{Thermal Management and Communication Virtualization for Reliability Optimization in MPSoCs}},
  booktitle = {Embedded},
  year = {2020},
  doi = {10.1007/978-3-030-52017-5\_8},
}
Figure