FT-EALU: Fault Tolerant Arithmetic and Logic Unit for Critical Embedded and Real time Systems

TL;DR

This paper introduces FT-EALU, a fault-tolerant arithmetic and logic unit for embedded systems that uses data diversification, weighted voting, and a design-time learning scheme to effectively mitigate transient and permanent faults.

Contribution

The paper presents a novel fault-tolerant approach combining data diversification, weighted voting, and a learning scheme to improve fault correction in embedded ALUs.

Findings

Corrects approximately 85% of single-bit permanent faults.

Corrects approximately 70% of double-bit permanent faults.

Demonstrates efficiency through experimental validation.

Abstract

In this paper, a fault-tolerant approach to mitigate transient and permanent faults of arithmetic and logic operations of embedded processors called FT-EALU is proposed. In this method, each operation is replicated in time and the derived final results are voted to generate the final output. To consider the effect of permanent faults, replicating identical operations in time is not sufficient, and diversifying the operands is required. To this aim in FT-EALU, we consider three distinct versions of input data and apply the target operation to them serially in time. To avoid high time overhead, we employ simple operators such as shift and swap to make an appropriate diversion in input data. Our proposed fault tolerance approach passes the replicated and diverse results to a novel weighted voter that is designed based on the reward/punishment strategy. For each version of execution, based on the proposed weighting approach a corresponding weight according to its correction capability confronting several faulty scenarios is defined. This weight defines the reliability of the result of each version of execution and determines its effect on the final result. The final result is generated bit by bit based on the weight of each execution and its computed result. These weights are determined statically through a design-time learning scheme according to applying several types of faults on various data bits. Based on the capability of execution versions on mitigating the permanent faults, positive or negative scores are assigned to them. These scores are integrated for several cases and normalized to derive the appropriate weight of each execution at bit level. Several experiments are performed to show the efficiency of our proposed approach and based on them, FT-EALU is capable of correcting about 84.93% and 69.71% of permanent injected faults on single and double bits of input data.