REAL: A Retention Error Aware LDPC Decoding Scheme to Improve NAND - PowerPoint PPT Presentation

REAL: A Retention Error Aware LDPC Decoding Scheme to Improve NAND Flash Read Performance Meng Zhang ∗ Fei Wu ∗ Xubin He † Ping Huang † Shunzhuo Wang ∗ Changsheng Xie ∗ ∗ Wuhan National Laboratory for Optoelectronics, Huazhong University of Science and Technology † Department of Electrical and Computer Engineering, Virginia Commonwealth University May 5, 2016 Zhang et al. (HUST & VCU) REAL May 5, 2016 1 / 22

Outline Introduction 1 REAL: Our Solution 2 Evaluation 3 Conclusion 4 Zhang et al. (HUST & VCU) REAL May 5, 2016 2 / 22

NAND Flash and Issues Description NAND flash gains popularity in various markets. High performance, low energy, small size Increasing capacity, decreasing per-bit price SLC MLC TLC Bits per cell 1 2 3 P/E Cycles 100,000 3,000 1,000 ECC capability ∼ 4 ∼ 24 ∼ more Source: White Paper “Solid State Drive Technology” from HP, 2013. Results Storage density increases Lifetime and data reliability reduces Zhang et al. (HUST & VCU) REAL May 5, 2016 3 / 22

Program/Erase (P/E) Cycles and Retention Error 1 P/E cycles and retention error (domain error) influences on data reliability : Figure: Rates of various types of errors as P/E cycles P/E cycles increase , retention time long , data reliability decreases 1 Figure Source: Cai et al [2] DATE 2012 Zhang et al. (HUST & VCU) REAL May 5, 2016 4 / 22

Error correction codes (ECC) (BCH, LDPC) 2 To ensure data reliability → ECC used BCH (prohibitively high cost, insufficient) Low density parity check ( LDPC ) codes 0 10 Decoding Failure Probability 5 10 0 10 BCH code LDPC code (hard decision sensing) LDPC code (six extra sensing levels) 5 10 0 02 0 018 0 016 0 014 0 012 0 01 0 008 0 006 0 004 Raw Bit Error Rate Figure: Decoding performance comparison 2 Figure Source: Zhao et al [3] FAST 2013 Zhang et al. (HUST & VCU) REAL May 5, 2016 5 / 22

LDPC 3 Directly using LDPC (without optimization, high compexity) introduces non-trivial overhead (a) Decoding latency increases (b) NAND flash read performance degrades How to solve the issue? (a) Optimization LDPC (b) Reduce decoding latency and improve NAND flash read performance 3 References: Zhao et al [5] MSST 2014 Zhang et al. (HUST & VCU) REAL May 5, 2016 6 / 22

Outline Introduction 1 REAL: Our Solution 2 Evaluation 3 Conclusion 4 Zhang et al. (HUST & VCU) REAL May 5, 2016 7 / 22

Retention Error Characteristic Charges leakage (theshold voltage shifts) 1.4 1.4 LSB/Lower page 1 1 0 0 1.2 1.2 MSB/Upper page 1 0 1 0 Probility dens ty function P(v) Probil ty density function P(v) REF 2 REF 3 REF 1 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 −6 −4 −2 0 2 4 6 −6 −4 −2 0 2 4 6 Vref1 Vref2 Vref3 Cell threshold voltage Vth(v) Cell threshold voltage Vth(v) (a) Theshold voltage shifts (red line) (b) Mapping information bits to voltage Numerical correlation → Two bits from the same cell affect each other with charges leakage. 00 → 01 01 → 10 01 → 11 10 → 11 46% 44% 5% 2% Source: Cai et al [2] DATE 2012 η Zhang et al. (HUST & VCU) REAL May 5, 2016 8 / 22

REAL We propose REAL scheme to incorporate the numerical-correlation of retention errors into the process of LDPC decoding to provide additional bits desicion information to reduce decoding latency and improve NAND flash read performance write read I/O data LDPC input data buffer encoder NAND flash array pages I/O data LDPC output data buffer decoder numerical correlation Figure: Numerical correlation is integrated into the decoding process. Zhang et al. (HUST & VCU) REAL May 5, 2016 9 / 22

How to use the numerical correlation? 4 make full use of the numerical correlation? (1) adjust LDPC codewords layout (2) obtain several observations based on the numerical correlation (3) translate these observations into mathematical models (4) the translated mathematical models are added into the decoding process. 4 Specific details can be found in our paper Zhang et al. (HUST & VCU) REAL May 5, 2016 10 / 22

LDPC Codewords Layout 1 to make a codeword contains two bits from the same MLC NAND flash cell 2 to provide the chance and environment that can effectively utilize the numerical correlation characteristic for the proposed REAL scheme Conventional LDPC-codewords ... ... Lower page ... ... Upper page NAND flash cell ... ... ... ... Lower page ... ... ... ... Upper page codewords • codewords • Proposed LDPC-codewords layout •• Zhang et al. (HUST & VCU) REAL May 5, 2016 11 / 22

Our Observations 5 Lower page 1 1 0 0 " $ # ! • • Upper page 1 0 1 0 Figure: Observation overview. � � � � � � 1 We obtain several valuable observations based on the numerical correlation characteristic of retention errors ( MLC NAND flash ) 2 Symbols 1 � 2 � 3 � 4 � 5 � 6 � correspond to the 6 observations respectively. 5 More details can be found in our paper Zhang et al. (HUST & VCU) REAL May 5, 2016 12 / 22

� � � � � � � � � � Work Process 6 Proposed LDPC-codewords layout � � ... codeword- � ... codeword- � ... ... bit node check node When decoding, the additional decoding information E j is provided 6 The translated mathematical models can be found in our paper Zhang et al. (HUST & VCU) REAL May 5, 2016 13 / 22

Outline Introduction 1 REAL: Our Solution 2 Evaluation 3 Conclusion 4 Zhang et al. (HUST & VCU) REAL May 5, 2016 14 / 22

Evaluation Methodology Simulation build MATLAB simulation environment utilize the AWGN (Additive White Gaussian Noise) channel to simulate the flash channel LDPC configuration Codeword Length (256 × 9)B Information Length (Page Size) 2KB Column Weight 4 Row Weight 36 Code Rate 8/9 Experiment steps construct the check matrix information bits are encoded by applying the Gauss elimination algorithm adding simulated retention error noise and decoding Zhang et al. (HUST & VCU) REAL May 5, 2016 15 / 22

Evaluation Results (Decoding Iterations) (a) Decoding iterations comparison: Probability Domain BP (PD-BP), Logarithm Domain Min-Sum (LD-MS) and REAL (b) SP/SNP represents signal power to retention noise power ratio 100 PD−BP 90 LD−MS REAL Average decoding iteration numbers 80 70 60 50 40 30 20 10 0 3.5 3.6 3.7 3.8 3.9 4 4.1 4.2 4.3 4.4 SP/SNP(dB) Figure: Comparison of decoding iterations at different SP/SNP Zhang et al. (HUST & VCU) REAL May 5, 2016 16 / 22

Decoding Failure Rate (1) The failure rate of the decoded codewords is lower than the baselines (2) If the SP/SNP continues decreasing, the baselines are not able to correct the LDPC codewords (REAL can be competent) 0.9 PD−BP 0.8 LD−MS REAL Decoded codewords failure rate 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 3.5 3.6 3.7 3.8 3.9 4 4.1 4.2 4.3 4.4 SP/SNP(dB) Figure: Comparison of decoding iterations at different SP/SNP Zhang et al. (HUST & VCU) REAL May 5, 2016 17 / 22

Decoding Latency and Read Performance of NAND Flash REAL reduces decoding latency by 26.44% and 33.05%. REAL improves NAND flash read performance. 50% Read performance improvement of NAND flash (%) The proposed REAL scheme over LD−MS The proposed REAL scheme over PD−BP 40% 30% 20% 10% 0% 4.5 4.4 4.3 4.2 4.1 4 3.9 3.8 3.7 3.6 3.5 SP/SNP(dB) (a) The reduced decoding latency with (b) The improvement in NAND flash REAL. read performance of our scheme. Zhang et al. (HUST & VCU) REAL May 5, 2016 18 / 22

Outline Introduction 1 REAL: Our Solution 2 Evaluation 3 Conclusion 4 Zhang et al. (HUST & VCU) REAL May 5, 2016 19 / 22

Conclusion In order to reduce the influences of NAND flash retention errors and improve the stored information reliability, we adopt the LDPC codes with optimized decoding performance. Retention errors of NAND flash cells have the characteristic of numerical-correlation that motivates us to effectively leverage the characteristic in the process of LDPC decoding in order to decrease the decoding latency and thus improve the NAND flash read performance. we propose the REAL scheme that accounts for the numerical-correlation characteristic into the LDPC decoding process, which can reduce the LDPC decoding latency and improve the read performance of NAND flash. Zhang et al. (HUST & VCU) REAL May 5, 2016 20 / 22

Acknowledgement We would like to thank Xiaosong Ma, our shepherd, and the anonymous reviewers for their valuable comments that greatly improved our paper. This research is sponsored by the National Natural Science Foundation of China, National Natural Science Foundation of Hubei Province, the Education Ministry of Hubei Province of China, the National High Technology Research and Development Program of China (863 Program), Key Laboratory of Data Storage System, and U.S. National Science Foundation. Zhang et al. (HUST & VCU) REAL May 5, 2016 21 / 22

The End Thanks ! Zhang et al. (HUST & VCU) REAL May 5, 2016 22 / 22