TEMPORISE: Extracting semantic representations of varied input executions for silent data corruption evaluation

Abstract

The continuous advancement of technology has led to increasingly complex computing systems, but it has also made them more susceptible to soft errors. Among the challenges posed by soft errors, silent data corruption (SDC) stands out as a particularly insidious threat, often occurring without warning. Estimating SDC probabilities for a program is a formidable task due to the diversity of inputs it can encounter, resulting in significant variations in these probabilities. This paper introduces TEMPORISE, a novel approach designed to tackle this challenge. TEMPORISE leverages the control data flow graph and calling context tree to represent the commonalities and distinctions between different input executions. The embeddings of these graphs are learned through structured graph attention network and AttrE2vec. These embeddings are then combined and input into a regression model to calculate SDC probabilities. The experiments demonstrate that TEMPORISE excels in predicting SDC probabilities, achieving a 78.4 % reduction in mean absolute error compared to vTRIDENT, the state-of-the-art baseline model. Moreover, TEMPORISE improves the rank correlation of SDC probabilities for various inputs by 11.4 % compared to vTRIDENT, indicating its superior ability to capture the relative ordering of SDC probabilities. In terms of computational efficiency, TEMPORISE boasts an impressive 91.3 % reduction in time cost compared to the traditional fault injection approach.