Soft errors are increasing in computer systems due to shrinking feature sizes. Soft errors can induce incorrect outputs, also called silent data corruption (SDC), which raises no warnings in the system and hence is difficult to detect. To prevent SDC effectively, protection techniques require a fine-grained profiling of SDC-prone instructions, which is often obtained by applying machine learning models. However, these models rely on handcrafted features, and lack the ability to reason about SDC propagation, which leads to an inferior SDC prediction performance. We propose a novel Graph Attention neTwork to Predict SDC-prone instructions (GATPS). The GATPS representation is a heterogeneous graph with different types of edges to represent various instruction relations. By stacking layers in which nodes are able to attend over their neighborhoods’ features, GATPS automatically captures the structural features that contribute to SDC propagation. The attention mechanism is applied to compute the importance values to the neighboring nodes, which quantifies the fault effect on the neighboring nodes. Moreover, the inductive model of GATPS can be applied to unseen programs without retraining, and it requires no fault injection information of the target program. Experiments revealed GATPS achieved a 34% higher F1 score compared to the baseline method and a 40-fold speedup compared to the fault injection approach.

由于特征尺寸的缩小，计算机系统中的软错误正在增加。软错误会导致不正确的输出，也称为静默数据损坏 (SDC)，它不会在系统中引发警告，因此难以检测。为了有效地防止 SDC，保护技术需要对易发生 SDC 的指令进行细粒度的分析，这通常是通过应用机器学习模型获得的。然而，这些模型依赖于手工制作的特征，缺乏对 SDC 传播的推理能力，导致 SDC 预测性能较差。我们提出了一种新颖的G raph A注意力网络T工作来预测SDC 倾向指令 (GATPS)。GATPS 表示是一个异构图，具有不同类型的边来表示各种指令关系。通过堆叠节点能够参与其邻域特征的层，GATPS 自动捕获有助于 SDC 传播的结构特征。注意力机制用于计算相邻节点的重要性值，量化了故障对相邻节点的影响。此外，GATPS 的归纳模型无需重新训练即可应用于未见过的程序，并且不需要目标程序的故障注入信息。实验表明，与基线方法相比，GATPS 的 F1 分数提高了 34%，与故障注入方法相比，速度提高了 40 倍。