GPU-Accelerated Parallel Mamba with Retrieval-Augmentation for Linear-Complexity Log Anomaly Detection

Log anomaly detection is a core task in AIOps for ensuring system reliability. With explosive growth in log volume, Transformer-based methods face severe computational bottlenecks on long sequences due to their quadratic complexity ≤ft(O≤ft(N2)). In this paper we propose RAMamba, a GPU-parallelized, linear-complexity framework for log anomaly detection. At its core RAMamba employs a selective state-space model (Mamba) that attains linear complexity (O(N)) and supports parallel associativescan operations, significantly improving throughput on large-scale log streams. In addition, we introduce a temporal retrievalaugmentation mechanism that retrieves historically similar temporal patterns in vector space to mitigate the difficulty of modeling rare anomalies from a single sequence. Quantitative evaluation shows that RAMamba achieves state-of-the-art performance on the HDFS, BGL and Thunderbird datasets, reaching an F1 score of 0.987 and a recall exceeding 99.7% on BGL. Experiments demonstrate that RAMamba maintains high detection accuracy while substantially reducing computational cost, making it highly attractive for industrial deployment.