Abstract:

Reliability prediction of GPGPU programs under soft error perturbations presents significant challenges， particularly when fault injection samples are limited. This paper proposes a novel method for constructing a high-accuracy reliability prediction model using a small number of fault injection samples. The approach introduces guiding rules to optimize fault injection experiments by reducing redundancy and enhancing the quality of collected samples. Furthermore， data augmentation techniques are employed to mitigate the challenges of insufficient training data， enabling effective model learning. Experimental validation on various benchmark programs demonstrates that the proposed method achieves an average prediction accuracy of 91.56% with only 300 fault injection samples. These results highlight the method's capability to reduce experimental overhead while maintaining high prediction performance.

Key words: computer architecture, general-purpose graphics processing units, reliability, soft error