EnergyAware Data Compression for MultiLevel Cell MLC Flash Memory

1
Energy-Aware Data Compression for Multi-Level
Cell (MLC) Flash Memory

• Yongsoo Joo, Youngjin Cho,
• Donghwa Shin, and Naehyuck Chang
• School of Computer Science Engineering,
• Seoul National University, Korea

2
Outline

• Motivation
• Programming energy variation of a MLC NOR flash
• Energy-aware data compression for MLCflash
  memories
• Energy characterization of MLC NOR flash
• Energy measurement
• Energy modeling
• Verification
• Energy-aware entropy coding
• Optimal prefix-free codes with unequal
  andfloating letter costs
• Energy-size tradeoff
• Conclusion

3
Motivation

• Multi-level cell (MLC) technology
• The large gap between the threshold voltage (VT)
  of the erased state and the programmed state
• Intermediate state using the gap of VT
• Increased storage capacity with the same
  technology

Programmed state
Programmed state
Erased state
Erased state
lt Threshold voltage distributions for (a) an SLC
flash and (b) a four-level MLC flash gt
4
Motivation

• SECM SNU Energy Characterizer for Memory devices
• Cycle-accurate energy measurement tool based on a
  switched capacitor method
• Demonstrated at DAC/SIGDA University Booth 2007

5
Motivation

• Programming energy variation of a MLC NOR flash
• Intel Strata Flash 28F256L18 (256Mbit, 1.8V)
• 4-level MLC NOR flash
• 32-word buffered programming mode

6
Motivation

• Traditional data compression
• Equal bit-pattern costs
• In terms of size and energy
• Minimum data size ? minimum energy consumption
• Goal minimize the compressed data size
• Energy-aware data compression for MLCflash
  memories
• Unequal bit-pattern costs
• In terms of energy consumption
• Minimum data size ? minimum energy consumption
• New goal find the compression method that
  minimizes the programming energy

7
Energy Characterization

• Programming energy modeling
• Experiment
• 32-word buffered programming mode (256 cells at a
  time)
• Change in the number of 00, 01, and 10(the
  remainder are filled with 11)

• Cost type
• Incremental cost the slope of the line drawn by
  each bit-pattern
• Common-mode cost the y-axis intercept of the
  line drawn by each bit-pattern

8
Energy Characterization

• Programming energy modeling
• Formulation of the programming energy model
• Energy coefficients

9
Energy Characterization

• Programming energy modeling
• Verification

10
Energy-Aware Entropy Coding

• Entropy coding with unequal letter costs
• Solutions to construct optimal prefix-free codes
• ILP by R. M. Karp
• Dynamic programming by M. J. Golin et al.
• PTAS (Polynomial-time approximation scheme)by M.
  J. Golin et al.
• Predetermined letter costs
• Entropy coding for MLC NOR flash memories
• No way to determine each bit-pattern costbefore
  programming
• Programming multiple cells at a time
• Existence of common mode costs
• New probabilistic approach
• Expected value of bit-pattern costs for
  givenbit-pattern probabilities

11
Energy-Aware Entropy Coding

• Energy and size estimation for given
• The expected energy consumption of programminga
  data unit of an MLC flash
• The normalized information quantity that is
  programmed in a data unit of MLC flash
• The derived cost function
• Normalized energy consumption to program the same
  amount of information

12
Energy-Aware Entropy Coding

• Problem statement
• Find that minimizes
• Solution space exploration
• Direct search method
• Result energy-optimal bit-pattern probability
  pair
• Energy-optimal bit-pattern costs
• For the given bit-pattern probability pair, the
  cost of each bit-pattern can be
  decided(C. E. Shannon)
• Construction of optimal prefix-free codes
• The ILP formulation by R. M. Karp

13
Energy-Aware Entropy Coding

• Solution space exploration results
• Intel 28F256L18

14
Energy-Size Tradeoff

• Maximum utilization of the storage and battery
  resources of mobile devices
• Encoding alternatives
• Flash storage becomes full ? size-optimal
  encoding
• Battery becomes empty ? energy-optimal encoding
• Intermediate solutions

15
Conclusion

• Detailed analysis of the programming energy of
  MLC NOR flash memories
• Energy-aware entropy coding for MLC NOR flash
  memories
• Energy-size tradeoff
• Maximum utilization of the battery and storage
  resource of mobile devices

16
Thank you!