Memory failure prediction using online learning

Abstract

Occurring frequently in datacenters, dynamic random access memory (DRAM) errors are the leading cause of the failures among various hardware components. DRAM failure analysis is one of the most important topics in hardware reliability, availability, and serviceability. Though with comprehensive studies of DRAM failure modes in prior work, a mechanism of predicting future failures on DRAM components is not available today. In this paper we address the problem of predicting the failures on micro-level DRAM components including cells, rows, and columns. A DRAM failure is the combined effect of the wear level of a DRAM fault and the implicit runtime context. Correctly predicting DRAM failures quantifies the impact to DRAM reliability and enables advanced error-prevention mechanisms such as efficient page retirement or dynamic substitution with spare DRAM components. We propose an online learning method, repeatedly taking the historical memory failure data of an individual server as the input to predict its failure occurrences in the near future. The learning algorithm embeds a kernel function to evaluate how well the current error observation follows certain previous observations in its failure history. It then performs the prediction by discovering the implicit patterns online. The algorithm strengthens failure confidence with the history of repeated errors from hard faults while washing out occasional errors from soft faults. It also adapts to the unobservable time-varying context by penalizing the change in failure observations across cycles. We further augment the algorithm with a mechanism which propagates failure confidence scores to nearby cells including the ones without error history. Empirical evaluation demonstrates that the failure prediction approach consistently outperforms the baseline methods based on historical error statistics.