SSD Characterization: From Energy … Characterization: From Energy Consumption's Perspective...

50
SSD Characterization: From Energy Consumption's Perspective Balgeun Yoo, Youjip Won Seokhei Cho, Sooyong Kang Hanyang University Jongmoo Choi Dankook University Sungroh Yoon Korea University HotStorage ‘11 June 14, 2011, Portland OR

Transcript of SSD Characterization: From Energy … Characterization: From Energy Consumption's Perspective...

SSD Characterization:

From Energy Consumption's Perspective

Balgeun Yoo, Youjip WonSeokhei Cho, Sooyong Kang

Hanyang University

Jongmoo Choi

Dankook University

Sungroh Yoon

Korea University

HotStorage ‘11

June 14, 2011, Portland OR

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Outline

Motivation

SSD Organization and Energy Consumption

Channels, Ways and Clusters

Case Study

Power Budget

Conclusion

2 /39

Motivation

3 /39

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Motivation

Hard disk

Sector layout: cylinder serpentine vs. surface serpentine vs. hybrid

serpentine

Number of zones

Degree of track skew

Disk scheduling algorithm

4 /39

Understanding of Internal Mechanism of Storage Device is very important!

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Characterizing HDD: Sector Layout

Jongmin Gim et al, ACM ToS 6, 2 (July 2010)

5 /39

Surface serpentine

Cylinder serpentineSpindleSpindle

Traditional SpindleSpindle SpindleSpindle

Zone rewind SpindleSpindle

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Characterizing Storage Devices

6 /39

HDD Characterization is via measuring Seek time and Rotational Latency.

Characterizing SSD… what do we use?...

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Motivation

What is available the number of channels

the number of chips/packages per channels

What is not available? Sector placement , Garbage collection algorithm

7 /39

How do we figure out the internals of SSD?

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Operations on NAND Flash Cell

8 /39

Erased (1)Programmed (0)

Substrate

Source Drain

Floating Gate

Control GateOxide Layer

- - -

18~35V

Substrate

Source Drain

Floating Gate

Control Gate

- - -

program erase

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Operations on NAND Flash Cell

9 /39

Youjip Won Hanyang University HotStorage ’11, Portland, OR

SSD Characterization

We will use “Energy Consumption”

10 /39

SSD Organization and Energy Consumption

11 /39

Youjip Won Hanyang University HotStorage ’11, Portland, OR

SSD Organization

12 /39

ControllerDRAM

I/OPhysical

Interface

Bus

Host Interface

NAND Flash Memory

Host

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Characteristics of SSD Behavior

13 /39

=Page Write

=Write Complete

Youjip Won Hanyang University HotStorage ’11, Portland, OR

IO Size vs. Energy Consumption

14 /39

Channels, Ways, and Clusters

15 /39

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Writing 4 pages: 1 Channel X 1 Way

16 /39

Cur

rent

(mA

)

time

P

=Page Write

C D

=Write Complete

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Way switch vs. Channel switch Delay

17 /39

Cur

rent

(mA

)

time

=Page Write

C D P

=Write Complete

Cur

rent

(mA

)

time

P

C D

Way Switch Channel Switch

Way Switch Channel Switch

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Writing 4 pages: 1 Channel X 2 Way

18 /39

Cur

rent

(mA

)

time

=Page Write

C D P

=Write Complete

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Writing 4 pages: 2 Channel X 1 Way

19 /39

Cur

rent

(mA

)

time

P

=Page Write

C D

=Write Complete

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Writing 4 pages: 2 Channel X 2 Way

20 /39

Cur

rent

(mA

)

time

P

=Page Write

C D

=Write Complete

Way Switch

Channel Switch

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Cluster

Cluster : Write Unit of SSD

21 /39

=Page Write

=

4KB Write 8KB Write

Case Studies

22 /39

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Measurement

Oscilloscope (TDS3032)

High resolution current probe(TCP202)

Current probe to power line(Vdd) of the

SSD

Sampling interval(10samples MA): 10usec

23 /39

Model Vendor Size Channels DRAM Size Package Type

X25M Intel 80GB 10 16MB 20 MLC

MXP SAMSUNG 128GB 8 128MB 16 MLC

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : Intel X25M

24 /39

Capacity 80GB

No. of Channels 10

Packages/Channel 2

Package 4 GB

NAND(4GB)

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : Intel X25M

IO Size: 4KB to 80KB

25 /39

100mA

17mA

Increase20 steps

Channel switch = 30 μsec

NAND programming: 17 mA

Simple Round Robin600 μsec

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : Intel X25M

4KB to 80KB Sequential Write

26 /39

= Page Write

0 2 3 4 5 6 7 8 91

10 121314151617181911

MUX

Channel Switch =30 μsec

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : Intel X25M

80KB to 160KB Write

27 /39

80 KByte

84 KByte

?

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : Intel X25M

28 /39

=1 Page Write

0 2 3 4 5 6 7 8 91

10 12131415161718191120

=Write CompleteMUX

time

P1P0

P2P3P4P5P6P7P8P9P10P11P12

P14P P13 P

P15P16

P19P18

P17

P20

Flo

w

time

Cur

rent

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : SAMSUNG MXP

29 /39

NAND(8GB)

Capacity 128GB

No. of Channels 8

Packages/Channel 2

Package 8 GB

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : SAMSUNG MXP

30 /39

Peak Current

100mA50mA

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : SAMSUNG MXP

31 /39

Same Peak Current

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : SAMSUNG MXP

32 /39

Same Peak Current

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : SAMSUNG MXP

33 /39

Cluster Size of MXP is 32KB

Same Peak Current

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Case Study : SAMSUNG MXP

What we still do not know of

Peak of 128 Kbyte < Peak Current of 124 KByte

34 /39

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Forth Coming Problem in Multi-channel SSD

35 /39

2008 2009 2010 2011

Read Write Capacity channel

Model Name Release

SATA3035 (Mtron) 2008.01

Vertex (OCZ) 2009.03

Vertex2 (OCZ) 2010.07

REVO Drive X2 (OCZ) 2011.01

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Forth Coming Problem in Multi-channel SSD

10 channel: peak 500 mA

16 Channel: peak 800 – 900 mA SSD is no long Green.

Further, excessive peak current can cause…

36 /39

supply voltage drop

ground bounce

signal noise

black-out

Etc…

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Power Budget

37 /39

Feasible Solutions

8 Channels X 1 Way

1 Channel X 8 Ways

Power Budget

Perfo

rmance B

ound

4 Channels X 2 Ways

2 Channels X 4 Ways

Power

I/O Latency

Feasible Region

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Summary

38 /39

32KB

35mA

350mA

Cluster Size : 4KB

programming: 17 mA

Peak Current : 500mA

Channel switch: 30 μsec

MXPX25M

Energy

Performance

Small Write

Large write

Standby current

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Summary

Energy Consumption is very good tool to characterizeSSD.

For larger number of channels, peak current will soon be a

significant issue.

We introduce the notion of Power Budge to resolve this

issue.

39 /39

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Related Works

Dongkun Shin et al, NVRAMOS 2010 Spring

Laura M. Grupp et al, MICRO 2009

41 /47

Applied different workloads(random, sequential, etc) to SSDs andMeasured the power consumption

Read

Program

Erase

Custom Board+ Flash Memory

Basic Operation

Power Consumption of Flash Memory Basic Operations

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Related Works

Euiseong Seo et al, HotPower’08

Vidyabhushan Mohan et al, Date ’10

42 /47

CACTI 5.3

developed a detailed power modelfor the NAND flash chip itself with CACTI 5.3

Applied different workloads(random, sequential, etc) to SSDswith different request sizes (varied the file systems)

Measured the Power Consumption

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Operations on NAND Flash Cell

43 /47

ModelRead (mA)

Write (mA)

Erase (mA)

Voltage(V)

MD332B (Intel) 20 20 20 3.3

K9XXG08XXM (Samsung) 25 25 25 3.3

MT29FXXG08CXXBB (Micron) 20 20 20 3.3

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Power Consumption of Storage Devices: System Boot

44 /47

stand-by current s

Active current of HDD

SSDs have various stand-by currents

Youjip Won Hanyang University HotStorage ’11, Portland, OR

FTL

FTL

45 /47

LPN: Logical Page Number PPN: Physical Page Number

LPN

PPN

0

12

1

11

2

10

3

1

4

3

5

7

Mapping Table

Flash Memory

Page 0 Page 8…

Block 0 Block 1 Block 2

Page FTL Block FTL

PBN: Physical Block Number

Hybrid FTL

Write toLPN = 5

Write toLPN = 3

LBN

PBN

0

5

1

1

2

7

Mapping Table

Flash Memory

Page 0 Page 8…

Block 0 Block 1 Block 2

Write toLPN = 5

Write toLPN = 3

LBN = 1Offset =

2

Page 0 Page 8

Block 0 Block 1 Block 2

Flash Memory

Write toLPN = 5

Write toLPN = 3

LBN = 1

LBN

PBN

0

5

1

1

2

7Block-Level

Mapping

LBN = 1Offset =

0

LPN

PPN

3

1

4 5

0Page-LevelMapping

Trivia in Measurement Methodology

46 /47

Youjip Won Hanyang University HotStorage ’11, Portland, OR

Trivia of Measurement Methodology

Sampling interval should be small enough.( Read/Program Latency)

Smoothing the data to filter out measurement noise.

47 /47

Smoothing