Evaluation Of Soft Error In Program Via Fusing Hardware Attributes

Soft error issue is a significant factor affecting the reliability of microprocessors.Evaluating soft errors and enhancing system reliability become important research topics.The Architectural Vulnerability Factor(AVF)represents the probability of soft errors ultimately resulting in visible errors in program output and is one of the important indicators for evaluating system reliability.The ACE analysis is often used for evaluating AVF.However,the drawback of ACE analysis is the lack of detailed information on microprocessor structure,and requiring manual classification of ACE and un-ACE bits for each program.When the hardware platform of the execution is changed,AVF can only be recalculated.In addition,AVF cannot directly reflect the probability of Silent Data Corruption(SDC),which has the greatest impact on users and is the most difficult to detect.This paper proposes a deep learning model that fusion hardware attributes to learn the propagation of soft errors from hardware to software.Based on the SDC propagation path,the Output Vulnerability Factor(OVF)is proposed to evaluate SDC-causing variables.The effectiveness of the method is validated through experiments.The innovative work of this paper is as follows:1)To capture the propagation of soft errors from the hardware layer to the software layer,this paper proposes a program soft error evaluation method of fusion hardware attribute.Using a multi-relation convolutional neural network to fuse the vector representations of hardware structure graphs,program dependency graphs,and dynamic dependency graphs.The prediction task of AVF is converted into a regression task of graph neural networks and applied to AVF prediction in various hardware environments.The experimental results show that the mean absolute error of the method proposed in this paper is reduced by 0.031 compared with the software-level vulnerability assessment methods.Our method can also perform inductive learning and provide relatively accurate prediction values for unseen hardware environments.2)To detect SDC,this paper proposes an SDC vulnerability detection method based on output vulnerability factors(OVF).The calculation of OVF is based on an enhanced Dynamic Dependence Graph(eDDG),a proposed instruction-level fault propagation model.It can filter out the edges representing the identified crash propagation path and perform a backward traversal of the eDDG to obtain SDC propagation path.Further,fault masking probability is estimated for the edges referring to value comparison and logistic operation.the SDC vulnerability of the variable is ranked based on the magnitude of the OVF.Experimental results show that the proposed method achieves the SDC detection rate of65.0% when detecting the top 10% high OVF variables,which is improved by 12-21%compared with previous methods.