Abstract
Power-constrained fault-tolerance has emerged as a key challenge in deep sub-micron technology. Multi-/many-core chips can support different hardening modes considering variants of redundant multithreading (RMT). In dark silicon chips, the maximum number of cores that can simultaneously be powered-on is constrained by the thermal design power (TDP). This chapter presents a system-level power–reliability management technique that jointly considers multiple hardening modes at the software and hardware levels, offering distinct power, reliability, and performance properties. A framework for system-level optimization is introduced to address power–reliability–performance management for many-core processors under varying user constraints.
Introduction: The Dark Silicon Challenge
As transistor sizes shrink into the deep sub-micron range, we face the "Dark Silicon" problem: while we can integrate hundreds of cores onto a single chip, we cannot power them all simultaneously without exceeding the Thermal Design Power (TDP). Systems must either power-gate cores (leaving them "dark") or run them at lower voltages and frequencies ("dim silicon").
However, reducing voltage increases the rate of transient faults (soft errors). Furthermore, manufacturing process variations mean that no two cores on the same chip are identical; some are naturally faster, while others leak more power. As shown in Figure 2a, these variations create a heterogeneous landscape that must be managed to maintain system reliability.
The PAFT Methodology
The proposed PAFT (Power-Aware Fault-Tolerance) technique employs a cross-layer approach to navigate these constraints. It bridges the gap between application-level requirements and architectural constraints, as illustrated in Figure 1.
PAFT operates in two distinct phases:
- Design-Time Code Selection: A reliability-aware compiler generates multiple versions of each application, ranging from high-performance to high-reliability (software-hardened) variants.
- Run-Time Management: A lightweight heuristic replaces traditional, computationally expensive Integer Linear Programming (ILP) solvers. This manager dynamically selects the hardening mode (e.g., SEDR, DMRR, or TMR), maps tasks to specific cores based on their unique physical profiles, and scales voltage/frequency (V-f) levels to stay within the TDP budget.
Key Insights and Results
The core innovation of PAFT lies in treating core heterogeneity as an opportunity rather than a drawback. By intelligently mapping timing-critical tasks to "fast" cores and power-hungry tasks to "efficient" cores, the system optimizes global reliability without missing deadlines.
Experimental evaluations using the MiBench suite on a simulated 64-core LEON3 processor yielded several key findings:
- Decision Speed: The PAFT heuristic is up to 1680x faster than an ILP solver, making it suitable for real-time operating systems.
- Reliability Improvement: PAFT achieved at least one order of magnitude higher reliability compared to single-layer techniques like dsReliM or DRVS.
- Low Overhead: The management overhead resulted in less than a 3% increase in total execution time.
- TDP Compliance: The system successfully maintained reliability even when 60% of the chip was forced to remain "dark" due to power constraints.
Conclusion
PAFT demonstrates that a synergistic, cross-layer approach is essential for managing the complex trade-offs between power, performance, and reliability in the dark silicon era. By leveraging both software flexibility and hardware redundancy, it provides a robust solution for next-generation many-core embedded systems.
Summary updated