Technical Papers Archive

Temporal Modeling of Soft Error Propagation Using Dynamic Bayesian Networks


Authors: Ahmer Jamil and Md Hasanur Rahman (University of Florida), Bingzhe Li (The University of Texas at Dallas), and Xiaoyi Lu and Guanpeng Li (University of Florida)

Abstract: With the ever-shrinking size of transistors, modern devices have become increasingly susceptible to soft errors that can lead to silent data corruption (SDC), seriously compromising system reliability. Understanding how these corruptions propagate through a program's execution is essential for accurate resilience analysis, yet existing analytical models rely on static assumptions that miss temporal behavior, while fault-injection studies reveal only final outcomes without exposing the intermediate propagation paths. This paper introduces Hydra, a temporal error-propagation framework that constructs a dynamic Bayesian network (DBN) from the program execution trace to model how corrupted values traverse dataflow, control-flow, and memory-dependency transitions. Hydra captures masking, reactivation, multi-step propagation, and long-range dependencies, enabling estimation of program-level SDC probabilities without repeated injections and revealing interpretable propagation pathways. Across a diverse benchmark suite, Hydra's predicted SDC probabilities closely match fault-injection results, achieving a mean absolute error of 2.2 percentage points, while substantially reducing computational cost.


Presentation: file

Back to Technical Papers Archive Listing