CE2Graph: A Diagnosis-Oriented Spatio-Temporal Graph Framework for DRAM Failure Prediction

Abstract

Dynamic random access memory (DRAM) failures, especially uncorrectable error (UE) and correctable-error (CE) storms, represent a major cause of node outages in modern data centers. Existing DRAM failure prediction methods mainly rely on flattened log features or handcrafted spatio-temporal statistics, which are insufficient for capturing the structured propagation of CEs across the hierarchical physical organization of DRAM. This paper proposes CE2Graph, a diagnosis-oriented spatio-temporal graph framework for DIMM-level DRAM failure prediction from correctable-error logs. CE2Graph transforms irregular CE logs into a graph representation, where nodes encode CE events along with physical-location metadata and error-bit patterns (e.g., DQs and Beats), while edges capture temporal adjacency and structural relations induced by shared DRAM components such as rows, columns, and banks. This representation enables explicit modeling of fault locality, temporal evolution, and bursty CE-storm behaviors under weakened error-correcting code (ECC) protection. Based on CE2Graph, we further develop a gated graph neural network (GGNN) with max-based neighborhood aggregation to emphasize severe localized hardfault patterns. Experimental results on large-scale real-world DRAM error and failure logs show that the proposed method consistently outperforms baseline models, including XGBoost and LSTM, in predicting future DRAM failures. These results suggest that graph-based spatio-temporal modeling is a promising direction for memory diagnosis and failure prediction.