How Much Can Data Compressibility Help to Improve NAND Flash Memory Lifetime?

Transcription

1 How Much Can Data Compressibility Help to Improve NAND Flash Memory Lifetime? Jiangpeng Li *, Kai Zhao *, Xuebin Zhang *, Jun Ma, Ming Zhao, and Tong Zhang * * Rensselaer Polytechnic Institute Shanghai Jiao Tong University Florida International University

2 NAND Flash Memory Bit cost reduction Price/GB Lifetime 2

3 Damage of NAND Flash memory cell q Damage to the NAND Flash memory cell Floating gate Oxide Electrons Oxide layer becomes thinner Number of electrons held in floating gate is reduced P/E cycling causes Charge trap : Damage to memory cell q Threshold voltage distribution overlap 3

4 Method to improve the lifetime Solid State drive Data Program NAND Read Data Log-Structured File system Damage causes limited lifetime Flash Translation Layer Data manipulation data compression Error Correction Coding Data compressibility NAND Uncompressed data Compressed data NAND Saved space Lifetime 4

5 Data compression in NAND Flash Data compressibility :α? Lifetime extends: 1/α q Unused space in one NAND Flash page One sector data Physical NAND Flash page data compression q Impact of compression ratio variance Unused space Minimize{damage} Compressed data Unused space Type-III Lower page Upper page wordline uncompressed data compressed data Type-I Type-II MLC NAND Flash wordline 5

6 Content-dependent damage characteristics Floating gate Oxide Electrons Floating gate State Lower page Upper page Erase 1 1 P1 P2 P V th Ø Electrons quantity differentiate the damage of data patterns Ø NAND Flash memory experiences content-dependent memory damage Ø Damage of data pattern: pattern 11 < pattern 10 < pattern 00 < pattern 01 6

7 Content-dependent damage: test result Raw BER Pattern content 11 Pattern content 10 Pattern content 00 Pattern content 01 Random pattern content P/E cycles x nm MLC NAND Flash chip 7

8 Unused space filling strategies Lower page Compressed data l head Unused space Type- III S (l) : Unused space in lower page S (u) : Unused space in upper page b l : lower page bit in memory cell k b u : upper page bit in memory cell k Upper page Unused space filling rule: Type- III Type- I Type- II b () l Ø If and (Type-III) l S b b Set ; l u = b = 1 u S ( u) Increase pattern 11 and 10 ; Decrease random data pattern and 01 Unidirectional data layout (UD) b () l Ø If and (Type-II-a) b l S b = 1 Set ; l () l b u S ( u) Ø If and (Type-II-b) l S b Set. l = b u b u S ( u) 8

9 Bidirectional data layout q Bidirectional data layout (BD): To reduce the percentage of random data pattern (Type-I) Compressed data Unused space l head Lower page Upper page Type- II Type- III Type- I Type- II 9

10 Conditional data exchange l head Type- III Compressed data Unused space l head Type- III Lower page Upper page Damage is reduced Type- III Type- I Type- II Type- III Type- I Type- II Type-II: pattern 11 or 00 Type-II: pattern 11 or 10 damage of 10 < damage of 00 q Conditional data exchange: exchange the compressed lower page data with upper page data to ensure the compressed lower page data has larger unused space May introduce extra overhead for FTL 10

11 Explicit compression and implicit compression q Conventional compressed data storage : Explicit compression Ø Complicate FTL/file system design q Implicit compression : A transparent data compression stored method Compressed data sector Unused space Compared with non-compression storage, the number of data sector in one Flash page is not increased Ø Transparent to FTL and file system, simplify system design Ø Sacrifice data compressibility and damage reduction 11

12 Damage factor of data pattern q 20nm MLC NAND Flash test results BER tolerance limitation η η 11 max r max q Damage Factor ρ i : Measure the damage of data pattern i η ρi = η ( r) max i max 12

13 Mathematical model to evaluate Flash lifetime : Model input : Model output Compressed data length in one page : ( e) C S Data compressibility (mean, variance) Mathematical model Lower page Upper page Compressed data Type- III Type- I Type- II Unused space Type- III After t P/E cycles, NAND Flash memory device survival probability Mathematical model Block survival probability Distribution of Mathematical model Damage factor of data patterns Distribution of memory cell damage per P/E Mathematical model ( e) C S Wordline survival probability Mathematical model NAND Flash physical parameters 13

14 Simulation results q Selected data compressibility (LZ77 compression algorithm, 4kB data sector length) 14

15 Simulation results q Lifetime improvement by data compression storage techniques DLL Text EXE XML 15

16 Impact of data compression ratio mean q Compression ratio standard deviation fixes at 0.01 Lifetime gain exp+ud exp+bd exp+udc exp+bdc imp+ud imp+bd imp+udc imp+bdc Compression ratio mean 16

17 Impact of data compression deviation q Compression ratio mean fixes at 0.5 Lifetime gain exp+ud exp+bd exp+udc exp+bdc imp+ud imp+bd imp+udc imp+bdc Compression ratio variance 17

18 Conclusion q Research on the employing data compressibility to reduce NAND Flash physical damage, in order to improve memory lifetime Ø Ø Ø Ø Content-dependent damage feature of MLC NAND Flash to reduce damage from data program and erase. A set of design strategies to exploit unused space from data compression in order to minimize the overall memory damage. File system/ftl-friendly implicit compression as a complementary of conventional explicit compression. Mathematical model to evaluate NAND Flash memory lifetime based on the data compressibility. 18