NAND flash memory has become increasingly popular in various computing systems. Although NAND flash memory offers attractive performance, it suffers limited operable programming and erasing cycles. To improve the reliability of flash-based systems, previous works introduce machine learning models to predict flash lifetime. These works generally focus on improving prediction accuracy but present little research about the resources required for flash lifetime prediction. In application scenarios, the overheads and the frequency of lifetime predictions are important for storage systems. Excessive prediction actions would lead to unnecessary resource consumption. For building an efficient storage system, resource requirements need to be taken into consideration when designing flash lifetime prediction schemes. In this paper, we propose adaptive lifetime prediction techniques (ADLPT) that minimize redundant prediction operations by exploiting reliability variation. To explore reliability variation, we investigate the error distribution of different 3D flash chips. Based on the investigation, a prediction judgment method is presented. The method identifies the necessary prediction by detecting the variation of erase duration and raw bit errors. Furthermore, we provide a method to improve the performance of the static model. The experimental result shows that our approach can reduce about 90% of redundant predictions with over 0.8 F1-Score.

NAND flash memory has gained popularity in a wide variety of digital storage systems. Although with excellent performance, NAND flash memory suffers various reliability problems. In recent years, researchers try to predict flash failure by using machine-learning models. However, the application of machine-learning based failure prediction method faces the following problems: imbalance between robustness and portability. When applying on different flash chips, the performance of prediction model degrades with the variation of error characteristics. In order to adapt to the variation, the machine-learning model needs to be re-built to ensure performance of failure prediction. The overheads of re-building model result in challenges when adjusting prediction model to adapt to the variation of error characteristics. To overcome these challenges, we present LightWarner, an easily applicable predictor based on model-free Reinforcement learning algorithms. LightWarner learns error characteristics dynamically during flash lifetime without pre-training. We evaluate the performance of LightWarner on six types of 3D flash chips. The evaluation result shows that LightWarner achieves over 93% F1 score on different flash chips, which is about 10% higher than supervised machine learning methods. And LightWarner can adapt to the variation of error characteristics with low migration costs.

Wear-out has been a critical reliability problem in NAND flash memory. As executing repeated program and erase operations on the NAND flash chips, the number of errors increases and ultimately exceeds the ECC capability. In previous work, error characteristics of flash wear-out are observed by endurance tests on a single type of NAND flash memory. We wonder if the experimental results cover the entire error characteristics of NAND flash memory. In this paper, we tested more than 20 types of NAND flash chips with different vendors and structures and presented an overlook of test results. Through the test results, we found an unexpected error-explosion phenomenon that errors of flash blocks first increase over several cycles and then reach a high value without warning. We analyzed the features of the error-explosion and explored its influence on operation time. And we propose an error-explosion prediction scheme to find the blocks that will occur an error-explosion in the next 1000 P/E cycles. The block identifying operation is realized by the machine-learning model. The performance of six machine-learning methods is compared. The results demonstrate that the Decision Trees and Bagged Classification Trees have the best accuracy.

In Solid State Drives, flash management techniques such as wear-leveling and refresh usually assume NAND flash memories have the same endurance value. However, the actual endurance values differ from blocks to blocks. This reliability difference is introduced by process-variation during flash fabrication. In recent years, for improving flash management techniques, various works have been done on the reliability variation of 2D flash memory. As 2D NAND transmitted to 3D NAND flash, the vertical structure and multi-layer stacking changed the effect of previously known reliability problems. In this paper, we are first to characterize the process-variation effects on 3D TLC flash reliability. The characterization includes two parts: endurance variation and error feature variation. Second, we propose an adaptive error prediction scheme to mitigate the process-variation effects. This scheme uses the machine-learning model to realize the error prediction operation. We also discuss the implications of this scheme on main flash management techniques.