- 學位論文
針對SSD應用之3D三層單元NAND快閃記憶體之錯誤預測模型
Error Prediction Model of 3D TLC NAND Flash for SSD Application
摘要
NAND 快閃記憶體最近在數據存儲中得到了廣泛的應用。不幸的是，快閃記憶體塊存在一些可靠性問題。如果我們繼續對快閃記憶體塊進行編程和擦除，可靠性問題將變得更加嚴重，最終可能會出現一些錯誤，例如大量糾錯碼（ECC）失敗甚至數據丟失。為了防止這種情況的發生，我們使用機器學習技術來構建預測模型，以預測芯片內在特定原始誤碼率(RBER) 下塊的剩餘壽命。為了驗證預測的有效性，我們使用SSDsim 並運行內置的磨損均衡算法來比較在我們自己創建的基線、預測和雙約束實驗下的失敗塊數和浪費的P/E 週期數. 我們發現，與傳統方法相比，基於預測的解決方案確實可以提高可靠性。結果表明，預測和雙約束實驗中的失敗塊數少於基線實驗。並且隨著加入到預測結果中的偏差發生變化，預測實驗中浪費的P/E 週期數最終會少於基線實驗。由於更嚴格的限制，雙約束實驗浪費的P/E 週期數仍然更高。
並列摘要
NAND flash has been widely used in data storage recently. Unfortunately, flash blocks suffer some reliability issues. For example, repeated programming and erasure of flash blocks will eventually lead to error-correcting code (ECC) failures and even data loss. We use machine learning technology to build a prediction model to predict the remaining lifetime for blocks at the specific raw bit error rate (RBER) within the chip to prevent the situation. To verify the effectiveness of the prediction, we use SSDsim and run the built-in wear-leveling algorithm to compare the number of failed blocks and the number of wasted P/E cycles under the baseline, prediction and dual constraint experiments are created by ourselves. We find that compared with traditional methods, prediction-based solutions can indeed improve reliability. The results show fewer failed blocks in the prediction and dual constraint experiments than in the baseline experiment. Additionally, as biases were added to the prediction results changes, the number of wasted P/E cycles in the prediction experiment will eventually become less than the baseline experiment. The number of wasted P/E cycles of the dual constraint experiment is still higher due to stricter restrictions.