Instruction-level reliability improvement for embedded systems

With the increasing number of applications in embedded computing systems, it became indispensable for the system designers to consider multiple objectives including power, performance, and reliability. Among these, reliability is a bigger constraint for safety critical applications. For example, fault tolerance of transportation systems has become very critical with the use of many embedded on-board devices. There are many techniques proposed in the past decade to increase the fault tolerance of such systems. However, many of these techniques come with a significant overhead, which make them infeasible in most of the embedded execution scenarios. Motivated by this observation, our main contribution in this paper is to propose and evaluate an instruction criticality based reliable source code generation algorithm. Specifically, we propose an instruction ranking formula based on our detailed fault injection experiments. We use instruction rankings along with the overhead tolerance limits and generate a source code with increased fault tolerance. The primary goal behind this work is to improve reliability of an application while keeping the performance effects minimal. We apply state-of-the-art reliability techniques to evaluate our approach on a set of benchmarks. Our experimental results show that, the proposed approach achieves up to 8% decrease in error rates with only 10% performance overhead. The error rates further decrease with higher overhead tolerances.