SDC Error Detection by Exploring the Importance of Instruction Features

Abstract

With the continuous improvement of the integration of semiconductor chips, it has brought great challenges to the reliability and safety of the system. Among them, Silent Data Corruption (SDC), as one of the most harmful issues, is difficult to be detected due to its concealment. For the SDC error detection, literatures just focus on fault injection or program analysis, but ignore the relationship of instruction features on the SDC errors. To this end, we propose a SDC error detection model by analyzing the instruction feature importance to the SDC vulnerability (SDIFI) and design the vulnerability prediction method. For the SDC error detection, specifically, we first analyze the correlation between different instruction features and the SDC vulnerability, and characterize the importance of these features. Second, we propose a SDC error detection model based on the SDC vulnerability prediction by an improved LightGBM model, as well as detecting SDC errors by selective redundancy of high SDC vulnerability instructions. Experimental results on the Mibench benchmarks show that our method has better detection accuracy with low overhead.