Data Distribution Algorithms for Reliable Parallel Storage on

Flash based memoryParallel and distributed storage based on Flash memories

Discussion and Summary

Data Distribution Algorithms for ReliableParallel Storage on Flash Memories

Kathrin Peter

Zuse Institute Berlin

November 2008, MEMICS Workshop

Kathrin Peter Reliable Parallel Storage on Flash Memories



Motivation

Nonvolatile storage

Flash memory -Invented by Dr. FujioMasuoka 1984

Type of EEPROM

Usage in a RAID-likeconfiguration instead ofhard disk drives

Flash memory based storage: Alternative to hard diskdrives?




Outline

1 Flash based memory

2 Parallel and distributed storage based on Flash memories

3 Discussion and Summary




Application of Flash memories




Pros and Cons

Less power consumption

Higher access rates (in some cases)

Uniform access time for randomaccess - no seeks

Robustness (extreme temperatures,vibration, shock)

Price

Limited erase cycles

Flash management

Model 2.5” SATA 3.0 Gbps SSD 2.5” SATA 3.0Gbps HDDMechanism type Solid NAND flash based Magnetic rotating platters

Density 64 GByte 80 GByteWeight 73 g 365g

Active Power consumption 1 W 3.86 WOperating temperature 0°C-70°C 5°C-55°C

Acoustic Noise None 0.3 dBEndurance MTBF > 2M hours MTBF <0.7M hours

Av. access time 0.1 msec 17 msecRead performance 100 MB/s 34 MB/sWrite performance 80 MB/s 34 MB/s




Limited erase cycles and flash management

Floating gate

Control gate

Sourcen+

Drainn+

P−Substrate

layersInsulating oxide

Transistor

.. .

NAND − flash

blockpage pagepage page

block=erase unit

Memory cell: floating gatetransistor

Retention and Endurance

Typical page size: 16896 B(=2 KB+64 B spare)

Typical block size:64 pages = 128 KB




Mapping

Problem: How to map logical blocks to flash adresses?Disadvantage of linear mapping (one-to-one mapping):

Frequently-used erase units wear outIdentity mapping requires lots of copying to and from RAM(fixed block size). Example:

data data data dataflashmemory

RAM

copy to RAM

data data data datamodify

copy back

data data datadata

data data datadata

mod.

mod.erase

Solution: Sophisticated block-to-flash mapping andmoving around blocks: wear leveling, garbage collection




Mapping - example

Mapping of virtual block 5 to physical block 0, page 3

7

2

2 74

0123

4

5

flash memory virtual−to−logical example:

0

virtual block

= log. erase unit

page

page

= logical block = 1117

1

0 1 2

5

3

3

2 01

= phy. erase unit 0

erase unit map

page maps

logical−to−physical

5

page

block

1

2

3

0

Algorithms and Data Structures for Flash Memories, Gal, Toledo, 2004




Wear leveling - example

...

hot pool cold pool

olde

r

stop

agi

ng

order inpools

flash memory

On Efficient Wear Leveling for Large-Scale Flash-Memory Storage Systems, Li-Pin Chang, 2007




Garbage collection - example

Free space drops below a thresholdSelect blocks for reclamation

Copy all live pages to free pages somewhere elseChange mapping entry, update to new positionErase block and allocate pages as free

erase

erase block x erase block x

erase block y free page

invalid page

life page




Facts (local) wear leveling

Performance: Erase operation is slow

No in-place update

Controller: Efficient mapping and erase distribution




Flash memory in a RAID-like system

Aggregate higher bandwidth

Reliability (redundancy for fault tolerance)

Problem: uneven usage of flash memories, more writes toredundant blocks when data becomes updated

Update accumulatedaccess load

Data Redundancy

flash-RAIM: Redundant Array of Independent flashMemories




Uneven distribution of writes




flash-RAIM and wear leveling

There exist data distribution algorithms to place dataevenly on a single memory

Goal: long lifetime of a single memory

We work on even distribution of writes across all memoriesin a memory array

Goal: even usage of cells on all flash memories, reliabledata storage, high throughput

Method: global wear leveling




Data Distribution Algorithms

Staggered Striping

+ − − −++

Memory 1 Memory 2 Memory 3

Hierarchical dual−pool Algorithm

Different starting points of algorithmsExplicit placing of dataData movement after storing and local wear leveling




Discussion of Future work

Extension of the simulator to evaluate the global wearleveling algorithms

Parameter study (trade-off)Define metrics to compare algorithms

Lifetime of the flash-RAIMSpeedOverhead

Usage of traces




Summary

Use aggregated bandwidth and fault-tolerance

flash-RAIM

Simulator for evaluation

Next steps:Implementation and evaluation of the global wear levelingalgorithm




Erase cycles data/ redundancy memory

0

10

20

30

40

50

60

70

80

90

100

1.1968e+06 1.197e+06 1.1972e+06 1.1974e+06 1.1976e+06 1.1978e+06 1.198e+06 1.1982e+06 1.1984e+06 1.1986e+06 1.1988e+06

num

ber

of e

rase

cyc

les

physical erase block adress

distribution of access frequencies

Disk 1Disk 6


Documents

Data Distribution Algorithms for Reliable Parallel Storage on