Research On Data Retention Failure Prediction Method For 3D NAND Flash Memory

With the development of cloud computing and big data technology,China has entered a high-level digital era.The digital storage system has become an essential part of everyday life,ranging from daily consumption to national defense and the military.During data storage,storage devices suffer various data security threats.In these security threats,storage medium failure would result in data loss directly and affect data storage reliability.Among storage mediums,3D(Three-Dimensional)NAND flash memory has become one of the mainstream memories in the market with high storage density and low cost.Although with excellent performance,3D NAND flash has various reliability problems.During data storage,the information stored in the 3D NAND flash unit will be affected by a phenomenon like trap-assisted tunneling,resulting in data retention errors.Over time,the number of data retention errors continues to increase.When the system cannot correct the errors,the data in flash memory will become invalid.Therefore,the research of data retention failure prediction method for 3D NAND flash has great significance.In recent years,researchers have introduced machine learning algorithms to predict possible errors for avoiding failures in flash chips.However,the existing reliability prediction methods usually focus on improving prediction accuracy,and less study on the optimization of the resources required for failure prediction.In application scenarios,excessive prediction operations will lead to unnecessary resource consumption.In order to build an efficient storage system,both prediction accuracy and prediction cost need to be taken into consideration when designing failure prediction strategies.To solve these problems,we studied the data retention failure mechanism of 3D flash and proposed a process variation aware data retention failure prediction strategy,which reduces the prediction cost and redundant prediction operations by exploiting flash reliability variance.In order to explore the error characteristics of flash memory,we analyze the error distribution of different types of 3D NAND flash chips.Based on the error characteristics,we proposed a reliability variance sensing method,which identifies the necessity of prediction by detecting the change of erasure duration and raw bit error.A model improvement method based on the flash error feature is also proposed,which builds the model according to manufacturing technology and improves the adaptability of the prediction model through transfer learning.The results show that the proposed strategy can reduce about 90% of redundant prediction operations while ensuring 94.5% of prediction accuracy.When the operating environment has changed,the 3D NAND flash would show different error features.To adapt to the error feature variation,the machine learning model requires updating the model to ensure prediction performance.This needs continuous collection of new data sets.Frequent updating of prediction models may affect system performance.To solve this problem,we introduced an agent-based data retention failure prediction strategy for 3D NAND flash.Inspired by the reinforcement learning algorithm,agents generated by the prediction strategy optimize the prediction strategy to adapt to the change of error features at runtime.In the prediction strategy,two types of prediction agents have been implemented: an error increment prediction agent and an error fluctuation prediction agent.An agent group optimization method is also proposed,which reduces the number of prediction agents by building agents for the flash cells with similar error characteristics.The experimental results show that the agent-based prediction strategy can predict 3D NAND flash data retention failure with over 95% prediction accuracy under the variation of flash type,dwell time,and temperature.And it can conduct prediction and update operations within 10 μs.The agent group optimization method can reduce about 90%prediction agents when the wear stressing of flash memory cells is similar.According to the experimental results,the proposed failure prediction strategies can effectively predict the data retention failure in 3D NAND flash,achieving the research goal of this work.