Cost-efficient soft error protection for embedded microprocessors

Jason A. Blome, Shantanu Gupta, Shuguang Feng, Scott A. MahlkeCASES 2006 · DOI

TL;DR

As microprocessors shrink, they become more susceptible to soft errors. This paper introduces two cost-effective techniques, the Register Value Cache (RVC) and time-delayed shadow latches, to protect embedded processors from these transient faults. These methods offer significant fault coverage with minimal area and power overhead, making them suitable for safety-critical embedded applications.

What Problem Does This Paper Address?

Modern microprocessors are increasingly vulnerable to soft errors due to aggressive device scaling. These errors, often caused by energetic particle strikes, can flip bits in memory elements or logic gates, leading to transient incorrect machine states. While high-end servers have employed expensive fault tolerance techniques like modular redundancy, these are often prohibitive for the cost-sensitive embedded market. This paper addresses the need for efficient, low-overhead soft error protection specifically tailored for embedded microprocessors, where area and power are critical constraints.

Methods

The authors propose two novel techniques:

  • Register Value Cache (RVC): This mechanism protects the processor's register file, a significant source of software-visible errors. Unlike traditional Error Correcting Codes (ECC) which only protect the register state array, the RVC also covers faults in the combinational logic (read/write logic) of the register file. It maintains duplicate copies of recently used register values, providing nearly twice the fault coverage of ECC at a comparable or lower cost. The RVC uses a cyclic redundancy check (CRC) for error detection.
  • Time-delayed Shadow Latches: To protect against faults outside the register file, this technique strategically deploys transient pulse detectors. These detectors are placed at high fan-in sequential state elements, which are identified as critical nodes responsible for propagating errors. By using time-delayed shadow latches, these detectors can identify transient pulses and trigger pipeline flushes to correct errors. This approach allows for scalable fault coverage by adjusting the number of detectors.

The paper also details a comprehensive fault analysis framework using a Verilog model of an ARM926EJ-S embedded microprocessor. This framework injects faults into both sequential and combinational logic to analyze error propagation and masking behavior.

Datasets

The paper does not introduce new benchmarks but uses traces from the MediaBench benchmarks [9] (specifically rgb2yuv for image processing) to simulate the ARM926EJ-S processor and evaluate the proposed techniques.

Key Insights and Ideas

  • Register File Vulnerability: The register file, despite its relatively small area contribution, is the source of a disproportionately large number of software-visible errors. This highlights the critical need for robust register file protection.
  • ECC Limitations: Traditional ECC is insufficient for register files because it doesn't protect the crucial read/write combinational logic, which is a significant source of faults.
  • Combinational Logic's Insidiousness: Faults in combinational logic, while potentially having higher masking rates at the microarchitectural level, can cause more widespread corruption of architectural state and multi-bit errors compared to faults in sequential state elements.
  • Exploiting Fan-out: The significant fan-out of errors from combinational logic nodes can be leveraged to strategically place fault detectors, achieving high coverage with fewer detectors.
  • Scalability is Key: Both proposed techniques are designed to be scalable, allowing designers to tune fault coverage based on available area and power budgets.

Key Results

  • RVC vs. ECC: An RVC with as few as eight entries provides more than twice the fault coverage of an ECC-protected register file. Scaling to 14 entries achieves 99% fault coverage for register file faults with only a ~1% increase in area and power overhead compared to ECC.
  • Time-delayed Shadow Latches: A "coverage limit study" suggests that augmenting less than 30% of registers can achieve 99.9% fault coverage for faults outside the register file. In practice, augmenting ~25% of state elements achieved ~90% fault coverage.
  • Combined Approach: Applying both RVC and time-delayed shadow latches together can achieve approximately 84% fault coverage with less than 5.5% area overhead and about 14% power overhead.

Limitations and Future Work

  • Electrical Masking: The current analysis primarily focuses on logical and temporal masking, leaving electrical masking (attenuation of pulses by subsequent logic) for future work.
  • Automated Detector Placement: The paper mentions that an automated tool for placing time-delayed shadow latches is a subject of ongoing research and future work. The current evaluation uses a manual technique.
  • Probabilistic Assumptions: The RVC assumes the probability of multiple concurrent strikes corrupting both the RVC and the register file with the same incorrect value is negligible.
  • Workload Dependence: The fault analysis and evaluation were performed using an image processing benchmark (rgb2yuv). While representative, the behavior might vary with different workloads.

The authors suggest that their techniques can be further refined and integrated into automated design flows to provide cost-effective soft error protection for a wide range of embedded applications.

Summary updated

BibTeX
@inproceedings{blome2006cost,
  author = {Jason A. Blome and Shantanu Gupta and Shuguang Feng and Scott A. Mahlke},
  title = {{Cost-efficient soft error protection for embedded microprocessors}},
  booktitle = {CASES},
  year = {2006},
  doi = {10.1145/1176760.1176811},
}
Figure