Extending SSD Lifetimes with Disk-Based Write Caches · 2019. 2. 25. · - Contains desktop, SVN,...

Extending SSD Lifetimes with Disk-Based Write Caches

Gokul SoundararajanUniversity of Toronto

Vijayan PrabhakaranMahesh BalakrishnanTed WobberMicrosoft Research

1

Thursday, February 25, 2010

Solid-State Devices (SSDs)

2



‣ Replacing Hard Disk Drives (HDDs)

- Fast I/O reads

- Consume low power, no moving parts, and more reliable

2




- Fast I/O reads


‣ Limited write cycles

- Need to erase block before re-write

- High end SLC provide 100,000 erase cycles

- Mainstream MLC provides 5,000-10,000 erase cycles

2




- Fast I/O reads


‣ Limited write cycles

- Need to erase block before re-write

- High end SLC provide 100,000 erase cycles

- Mainstream MLC provides 5,000-10,000 erase cycles

‣ Starting to be used in laptops/desktops

- Contain write intensive workloads

2


SSD Write-lifetime

3


SSD Write-lifetime

3

Ideal Micron C200 Intel X-25M0

75

150

225

300

Wri

te L

ifeti

me

(TB

)Maximum amount of data that can be written to the

SSD


SSD Write-lifetime

3


75

150

225

300

Wri

te L

ifeti

me

(TB

)60GBx5000 =

300TB


SSD Write-lifetime

3


75

150

225

300

Wri

te L

ifeti

me

(TB

)

42TB (60GB drive)


SSD Write-lifetime

3


75

150

225

300

Wri

te L

ifeti

me

(TB

)

37TB (80GB drive)

20 GB/day*5 yrs


Griffin Hybrid Device

4

Write Cache SSD

Hybrid Device Controller

GriffinWrites

Reads Reads



4

Write Cache SSD


Griffin

MIGRATE

WritesReads Reads



4

Write Cache SSD


Griffin

MIGRATE

WritesReads Reads

Options



4

Write Cache SSD


Griffin

MIGRATE

WritesReads Reads

OptionsRAM



4

Write Cache SSD


Griffin

MIGRATE

WritesReads Reads

OptionsRAM

NVRAM



4

Write Cache SSD


Griffin

MIGRATE

WritesReads Reads

OptionsRAM

NVRAMSSD



4

Write Cache SSD


Griffin

MIGRATE

WritesReads Reads

OptionsRAM

NVRAMSSD

Hard Disk



4

Write Cache SSD


Griffin

MIGRATE

WritesReads Reads

Why Hard Disk?DurableCapacityCheap

Sequential I/O


Contributions

5


Contributions

‣ Characterized I/O patterns

- Found desktop/server traces contains many overwrites

- Caching overwrites reduces writes to SSD by 52% ideally

5


Contributions

‣ Characterized I/O patterns

- Found desktop/server traces contains many overwrites

- Caching overwrites reduces writes to SSD by 52% ideally

‣ Designed Griffin hybrid disk

- Uses a disk-based write cache

- Cache data on hard disk, periodically move it back to SSD

- Shows a 2x lifetime improvement ( < 5% HDD reads)

5


Outline

‣ Motivating Workload Characteristics

- Key features

‣Disk-based Write Caching

‣ Experimental Evaluation

‣Conclusions

6


I/O Workload Characterization

‣ Examined various I/O traces

- Desktops (Internal Microsoft traces)

- Servers (Narayanan et. al from MSR Cambridge)

- Use only the write-intensive traces

- Linux (Bhadkamkar et. al. from FIU)

- Contains desktop, SVN, and web server

‣ Trace descriptions

- Block I/Os collected below the filesystem buffer cache

- Multi-hour traces of 5 hours to 176 hours

- Between 209K to 543M I/O events per trace

7


Key Metrics

8


Key Metrics

‣Overwrites

- Consecutive writes to a block without an intervening read

8


Key Metrics

‣Overwrites


8

W: 100 W: 100 W: 100 W: 100 R: 100 W: 200 R: 200 W: 200 R: 200


Key Metrics

‣Overwrites


‣Write-after-Write (WAW) Times

- Time between two consecutive writes to the same block

8

W: 100 W: 100 W: 100 W: 100 R: 100 W: 200 R: 200 W: 200 R: 200

WAW


Key Metrics

‣Overwrites


‣Write-after-Write (WAW) Times

- Time between two consecutive writes to the same block

‣ Read-after-Write (RAW) Times

- Time between a write and a subsequent read to the same block

8

W: 100 W: 100 W: 100 W: 100 R: 100 W: 200 R: 200 W: 200 R: 200

RAWWAW


Why Overwrites? Lifetime of a Block

9



9

R: 100

Time ∞



9

W: 100 W: 100 W: 100 W: 100R: 100

Writeback to diskTime ∞



9

W: 100 W: 100 W: 100 W: 100 R: 100R: 100

Evicted from FS buffer

cache

Time ∞



9

W: 100 W: 100 W: 100 W: 100 R: 100R: 100

Evicted from FS buffer

cache

W: 100

Coalesced write to SSD

75% write savings

Time ∞


Top Overwritten Files in Desktops

10

Rank Filename

1 C:\Outlook.ost

2 C:\...\Search\...\Windows.edb

3 C:\$Bitmap

4 C:\Windows\Prefetch\Layout.ini

5 C:\Users\<name>\NTUSER.DAT

6 C:\$Mft


Ideal Write Savings: Remove Overwrites

11



11

0

25

50

75

100

Writ

e S

avin

gs (%

)

Desktops Servers Linux



11

0

25

50

75

100

Writ

e S

avin

gs (%

)


14%

94%



11

0

25

50

75

100

Writ

e S

avin

gs (%

) 52% AverageWrite Savings


14%

94%


WAW/RAW Time Intervals

12

WAW RAW

Cum

ulat

ive

Tim

e In

terv

al (%

)

Histogram Buckets (seconds)


0 10 30 60 120 300 600 900 3600 Inf0

25

50

75

100


12

WAW RAW

Cum

ulat

ive

Tim

e In

terv

al (%

)



0 10 30 60 120 300 600 900 3600 Inf0

25

50

75

100


12

Overwrites happen quickly

WAW RAW

Cum

ulat

ive

Tim

e In

terv

al (%

)



0 10 30 60 120 300 600 900 3600 Inf0

25

50

75

100


12

Written data is read

late

WAW RAW

Cum

ulat

ive

Tim

e In

terv

al (%

)



Summary of Observations

13



‣ Large fraction of overwrites

- Potential of 36% to 64% reduction for desktops

- As much as 94% in server workloads (web server)

- Linux: 62% in desktop and 81% in servers (web server)

13



‣ Large fraction of overwrites

- Potential of 36% to 64% reduction for desktops

- As much as 94% in server workloads (web server)

- Linux: 62% in desktop and 81% in servers (web server)

‣Overwrites happen quickly, reads after a long interval

- Over 50% of overwrites within 30 seconds

- Only 21% of written data read within 15 minutes

13


Outline



- Basic algorithm

- Performance tradeoffs

‣ Experimental Results

‣Conclusions

14


Griffin: Handling Writes

15

HDD SSD

Griffin

LOG: 1



15

HDD SSD

Griffin

W: 100LOG: 1



15

HDD SSD

Griffin

W: 100LOG: 1

Send to Hard Disk



15

HDD SSD

Griffin

LOG: 1LOG: 2 Block Log

100 1

200 2

Blockmap

Memory



15

HDD SSD

Griffin

LOG: 1LOG: 2LOG: 3 Block Log

100 1

200 2

Blockmap

Memory


Griffin: Handling Reads

16

HDD SSD

Griffin

LOG: 3 Block Log

100 1

200 2

Blockmap

Memory



16

HDD SSD

Griffin

R: 100LOG: 3 Block Log

100 1

200 2

Blockmap

Memory



16

HDD SSD

Griffin

R: 100LOG: 3 Block Log

100 1

200 2

Blockmap

Check Blockmap

Memory



16

HDD SSD

Griffin

LOG: 3R: 300

Block Log

100 1

200 2

Blockmap

Memory



16

HDD SSD

Griffin

LOG: 3 Block Log

100 1

200 2

Blockmap

Memory


Griffin: Data Migration

17

HDD SSD

Griffin

Block Log

100 1

200 2

Blockmap

Memory



17

HDD SSD

Griffin

Block Log

100 1

200 2

Blockmap

Migrate data to SSD

Memory



17

HDD SSD

Griffin

Block Log

100 1

200 2

Blockmap


Tradeoff: Write Savings vs. Read Penalty

18



18

What to cache?

How

long

to c

ache

?

All Writes

Long

Tim

e



18

Writ

e Sav

ings

What to cache?

How

long

to c

ache

?

All Writes

Long

Tim

e



18

Read P

enalt

yW

rite

Savin

gs

What to cache?

How

long

to c

ache

?

All Writes

Long

Tim

e



18

Read P

enalt

yW

rite

Savin

gs

What to cache?

How

long

to c

ache

?

All Writes

Long

Tim

e

What to cache?


What to cache?: Overwrite Distribution

19

0

25

50

75

100

0.01 0.1 1 10 100

Perc

enta

ge o

f Ove

rwrit

es

Percentage of Written Blocks


What to cache?: Overwrite Distribution

19

1% of the blocks receive

70% of overwrites

0

25

50

75

100

0.01 0.1 1 10 100

Perc

enta

ge o

f Ove

rwrit

es

Percentage of Written Blocks


Selective Write-Caching

20

HDD SSD

Griffin

LOG: 1



20

HDD SSD

Griffin

W: 100LOG: 1

Blockmap

Block Overwritten? Log

100 ✔



20

HDD SSD

Griffin

W: 100LOG: 1

Blockmap


100 ✔

Is heavily overwritten?



20

HDD SSD

Griffin

LOG: 1LOG: 2

Blockmap


100 ✔ 1



21

Read P

enalt

yW

rite

Savin

gs

What to cache?

How

long

to c

ache

?

All Writes

Long

Tim

e



21

Read P

enalt

yW

rite

Savin

gs

What to cache?

How

long

to c

ache

?

All Writes

Long

Tim

e

Ho

w lo

ng t

o c

ache

?


0

25

50

75

100

0 10 30 60 120 300 600 900 3600 Inf

How long to cache?: Time Intervals

22

Overwrites happen quickly

Written data is read

late

WAW RAW

Cum

ulat

ive

Tim

e In

terv

al (%

)



How long to cache?: Migration Triggers

23



‣ Timeout trigger

- Idea: Fires after a certain time elapses

- Could have high read penalty

23



‣ Timeout trigger



‣ Read-threshold trigger

- Idea: Fires if HDD reads exceeds threshold since last migration

- Bounds HDD read fraction

- Could have no migrations for a long time

23



‣ Timeout trigger



‣ Read-threshold trigger

- Idea: Fires if HDD reads exceeds threshold since last migration

- Bounds HDD read fraction

- Could have no migrations for a long time

‣Hybrid trigger

- Every 15 mins or read penalty > 5%

23


Outline



‣ Experimental Evaluation

- Caching Policies: What to cache?

- Migration Policies: How long to cache?

- Performance: Lifetime and Latency

‣Conclusions

24


0

10

20

30

40

50

60

70

80

90

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

SS

D w

rite

sav

ing

s

(per

cen

tag

e o

f to

tal

wri

tes)

Traces

Full CachingSelective Caching

What to Cache?: Write Savings

25


0

10

20

30

40

50

60

70

80

90

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

SS

D w

rite

sav

ing

s

(per

cen

tag

e o

f to

tal

wri

tes)

Traces



25

Migrated at the end of

trace


0

10

20

30

40

50

60

70

80

90

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

SS

D w

rite

sav

ing

s

(per

cen

tag

e o

f to

tal

wri

tes)

Traces



25


trace

25%


0

10

20

30

40

50

60

70

80

90

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

SS

D w

rite

sav

ing

s

(per

cen

tag

e o

f to

tal

wri

tes)

Traces



25


trace

51%

25%


0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

HD

D r

ead p

enal

ty

(per

cen

tag

e of

tota

l re

ads)

Traces


What to Cache?: Read Penalty

26


0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

HD

D r

ead p

enal

ty

(per

cen

tag

e of

tota

l re

ads)

Traces



26


trace


0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

HD

D r

ead p

enal

ty

(per

cen

tag

e of

tota

l re

ads)

Traces



26


trace

19%


How Long to Cache?: Write Savings

27

0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

SS

D w

rite

sav

ing

s

(per

cen

tag

e o

f to

tal

wri

tes)

Traces

Migration timeout 3600 sMigration timeout 1800 sMigration timeout 900 s



27

15min gives large savings

0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

SS

D w

rite

sav

ing

s

(per

cen

tag

e o

f to

tal

wri

tes)

Traces




27

15min gives large savings

1hr less additional

benefit

0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

SS

D w

rite

sav

ing

s

(per

cen

tag

e o

f to

tal

wri

tes)

Traces



How Long to Cache?: Read Penalty

28

0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

HD

D r

ead p

enal

ty

(per

cen

tag

e of

tota

l re

ads)

Traces



How Long to Cache?: Read Penalty

28

7%

0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C1

0

S−

SR

C2

2

S−

ST

G1

S−

WD

EV

2

HD

D r

ead p

enal

ty

(per

cen

tag

e of

tota

l re

ads)

Traces



0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C10

S−

SR

C22

S−

ST

G1

S−

WD

EV

2

SS

D b

lock

eras

ure

sav

ings

(%)

Traces

Savings with a simple FTLSavings with an ideal FTL

Performance Summary: Erasure Savings

29


0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C10

S−

SR

C22

S−

ST

G1

S−

WD

EV

2

SS

D b

lock

eras

ure

sav

ings

(%)

Traces



29

Direct block-level mapping


0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C10

S−

SR

C22

S−

ST

G1

S−

WD

EV

2

SS

D b

lock

eras

ure

sav

ings

(%)

Traces



29

Direct block-level mapping

Log structured page-level mapping


0

20

40

60

80

100

D−

1A

D−

1B

D−

1C

D−

1D

D−

2A

D−

2B

D−

2C

D−

2D

D−

3A

D−

3B

D−

3C

D−

3D

S−

EX

CH

S−

PR

XY

1

S−

SR

C10

S−

SR

C22

S−

ST

G1

S−

WD

EV

2

SS

D b

lock

eras

ure

sav

ings

(%)

Traces



29

> 2x


Performance Summary: I/O Latency

30

0

0.2

0.4

0.6

0.8

1

Total Read Write Total Read Write

Rel

ativ

e I/

O l

aten

cy

Workload T1 Workload T2

HDD 7.2KHDD 10K

MLCSLC



30

0

0.2

0.4

0.6

0.8

1


Rel

ativ

e I/

O l

aten

cy


HDD 7.2KHDD 10K

MLCSLC

Most I/O intensive segments (2hr) -

desktopThursday, February 25, 2010


30

0

0.2

0.4

0.6

0.8

1


Rel

ativ

e I/

O l

aten

cy


HDD 7.2KHDD 10K

MLCSLC

0.44



30

0

0.2

0.4

0.6

0.8

1


Rel

ativ

e I/

O l

aten

cy


HDD 7.2KHDD 10K

MLCSLC

0.44

0.15


Extending SSD Lifetimes with Disk-Based Write Caches · 2019. 2. 25. · - Contains desktop, SVN,...

Documents

Transcript of Extending SSD Lifetimes with Disk-Based Write Caches · 2019. 2. 25. · - Contains desktop, SVN,...