Abaci-finder: Linux kernel crash classification through stack trace similarity learning

Abstract

Developers often classify crashes by stack traces to analyze, locate and fix kernel bugs. Existing stack-trace-based crash classification approaches rely on string matching and statistical features, which ignore crash semantic contexts and cannot explore high-order correlations. Deep-learning-based approaches use crash embeddings and output end-to-end features for classification. However, they ignore kernel-specific information, which limits classification performance. Regarding these issues, we propose abaci-finder, a deep-learning-based classification framework specific to Linux kernel crashes. We first model the kernel stack trace as a stack frames sequence and then perform stack trace preprocessing. Then, we propose a vectorization method specific to kernel stack traces, called kstack2vec, to extract features with consideration for function semantics and kernel-specific offsets information. Finally, we exploit an attention-based BiLSTM neural network for classification, with consideration for both frame context and key frames in traces. The experiments on the real Linux kernel crash dataset indicate that abaci-finder outperforms existing methods of crash classification.