SmartDedup: Optimizing Deduplication for Resource ......Virtualized Infrastructures, Systems, &...

Virtualized Infrastructures, Systems, & Applications

SmartDedup: Optimizing Deduplication for Resource-constrained Devices

Qirui Yang Runyu Jin Ming [email protected], [email protected], [email protected]

Arizona State University

http://visa.lab.asu.edu

Resource Management on Edge and IoT

Data-intensive workloadso Multitude of sensorso Data-rich applicationso Multitasking systems

Limited on-device storageo Limited I/O performanceo Limited capacityo Limited endurance

2

• • •

• • •

Deduplication to the Rescue?

Well-known benefitso Eliminate redundant I/Os à improve performanceo Reduce flash writes à improve enduranceo Remove redundant data à improve utilization

Widely used in the cloudo Backup storage, primary storage, cache storage, …

But o Is there enough data duplication in device workloads?

o How to exploit it using limited resources on the device?

3SmartDedup

Smartphone Trace Analysis

Real-world device workloads are indeed I/O intensive

All workloads have a good level of duplicates

4SmartDedup

ID Daily I/Os(GB) Daily write

(GB)# of

months Duplication

ratio (%)

1 12.1 1.4 6 21.9

2 9.1 2.1 3 45.3

3 5.9 1 2.5 23.2

4 16.5 2.4 5 47.5

5 12.2 2.8 3 41.6

6 10.6 2.4 2.1 28.3

File system level traces collected from users from different countries

Trace: http://visa.lab.asu.edu/traces

More In-depth Analysis

5

Top 3 duplicate contributors for the traces

Trace1Trace2Trace3

Trace4Trace5Trace6

Five categories:

Resource files (res)

Database files (db)

Executables (exe)

Temporary files (tmp)

Multimedia (media)

More In-depth Analysis

Duplicate sources differ by devices

80% of duplicates are from files that are not entirely duplicate

System-level deduplication is necessary

6

Top 3 duplicate contributors for the traces

Trace1Trace2Trace3

Trace4Trace5Trace6

Challenges to Device Deduplication

Limited memoryo Deduplication relies on the in-memory fingerprint store for quickly

detecting duplicates

Limited storage performance and enduranceo Deduplication incurs additional metadata operations

Limited power and energy capacityo Fingerprinting is CPU intensive and draws quite a bit of power

7SmartDedup

SmartDedup—A Smart Deduplication Solution for Smart Devices

Cohesively designed two-level fingerprint storeso On-disk store complements the small in-memory store

Synergistically integrated hybrid deduplicationo Out-of-line deduplicates data skipped by in-line

Dynamically adapted deduplication o According to resource availability

and workload characteristics

8SmartDedup

In-memory Fingerprint Store

Store only important fingerprints

Organized by a prefix treeo Index groups of

fingerprints — small memory footprint

o Support the two-level fingerprint store

9SmartDedup

In-memoryFingerprint Store

First 6 bits0 1 2 63…

Second 6 bits

0 1 2 63…Second 6 bits

0 2 63…

leaf

prefix=0 prefix=129

…

…

FPa

FPb

PBNa

PBNbFPg PBNg

In-memoryfingerprint group

FPm

FPn

PBNm

PBNn…

FPq PBNq


…

leaf

1

On-disk Fingerprint Store

Maintain fingerprints that are not in the in-memory store

Share the same index with in-memory storeo Save memoryo Facilitate the fingerprint

migration

10SmartDedup


On-diskfingerprint groupPBNx

FPh

FPi

PBNh

PBNi…

PBNyFPm

FPn

PBNm

PBNn…

FPs PBNs

…



Second 6 bits


0 2 63…

leaf

prefix=0 prefix=129

…

…

FPa

FPb

PBNa

PBNbFPg PBNg


FPm

FPn

PBNm

PBNn…

FPq PBNq


…

PBNx

FPk PBNk

leaf

1

Fingerprints Migration

Fingerprints migrate between memory and disk on demand

In-memory store keeps the most recently used fingerprints

Fingerprints evicted by groups to save I/Os

11SmartDedup



FPh

FPi

PBNh

PBNi…

PBNyFPm

FPn

PBNm

PBNn…

FPs PBNs

…



Second 6 bits


0 2 63…

leaf

prefix=0 prefix=129

…

…


FPn

PBNm

PBNn…

FPq PBNq

FPm


…

PBNx

FPk PBNk

leaf

1

FPb

PBNa

PBNb…

FPt PBNt

FPa





12SmartDedup



FPh

FPi

PBNh

PBNi…

PBNyFPm

FPn

PBNm

PBNn…

FPs PBNs

…



Second 6 bits


0 2 63…

leaf

prefix=0 prefix=129

…

…


FPn

PBNm

PBNn…

FPq PBNq

FPm


…

PBNx

Free Space

FPk PBNk

leaf

1

FPb

PBNa

PBNb…

FPt PBNt

FPa





13SmartDedup



FPh

FPi

PBNh

PBNi…

PBNyFPm

FPn

PBNm

PBNn…

FPs PBNs

…



Second 6 bits


0 2 63…

leaf

prefix=0 prefix=129

…

…


FPn

PBNm

PBNn…

FPq PBNq

FPm


…

PBNx

Free SpaceFPk PBNk

leaf

1

FPb

PBNa

PBNb…

FPt PBNt

FPa

Hybrid Deduplication

In-line Deduplicationo Work in the I/O patho Immediately eliminate duplicate writes o Limited by the available resources on the device

Out-of-line Deduplicationo Work in backgroundo Slowly and thoroughly eliminate duplicate data

o Limited improvement on performance and endurance

14SmartDedup

Shared Fingerprint Index

Write Path

15SmartDedup

write

Skipped Buffer

In-memory FingerprintStore

Grp #1

Grp #2

On-disk FingerprintStore

Grp #1

Grp #2

…3

1 Fingerprinting2

Search in-memory store

Miss; pass it to out-of-line

4Search on-disk fingerprint store

4Search on-disk store

5Promote one fingerprint to in-memory store

If in-memory store full, evict a group of fingerprints to disk

6

Read Path

Native read path cannot utilize duplicate data in the page cacheo Page cache is indexed by logical block numbers (LBNs)o Cannot avoid reads to different LBNs if the same data is already in the page

cache

16

ReadLBNx

ReadLBNy

DATAxPBNx

DATAxLBNx

DATAxLBNy

DiskPage CacheMiss

Miss

PBNx

LBNx LBNyDuplicate Read

Optimized Read Path

Page Cache Indexo Map from PBNs to the corresponding pages in page cacheo Use this index to find data for reads with different LBNs

17SmartDedup

ReadLBNx

ReadLBNy

DATAxPBNx

DATAxLBNx

DiskPage CacheMiss

PBN-based IndexPBNx->DATAx

PBNx

LBNx LBNySave Read

Adaptive Deduplication

Based on resource availabilityo Keep CPU and disk utilization under 100%o Proportional to available battery lifeo Completely disable deduplication in low battery state

Based on the current duplication levelo Gradually reduce the rate if the duplication level is droppingo Quickly increase the rate if the duplication level is growing

18SmartDedup

Evaluation

19

Prototypeso EXT4 and F2FS (4KB fixed-size chunking)

Testing deviceso Nexus 5X and Raspberry Pi 3

Benchmarkso FIO—intensive I/O benchmarko DEDISbench—

workloads from real-world device imageso Real-world device traces

Baselineso Native EXT4 and F2FSo Dmdedup: block-level deduplication, flexible metadata backends [OLS 14’]o CAFTL: deduplication for resource-constrained FTL layer [FAST 11’]

Nexus 5X Raspberry Pi 3

CPU Qualcomm Snapdragon 808

Broadcom BCM2837

RAM 2GB 1GB

Storage 32GB eMMC 16GB SDHC UHS-1

OS Android Nougat Raspbian Stretch Lite

Kernel Linux 3.10 Linux 4.4

File System

EXT4 F2FS

FIO (on Nexus)

20SmartDedup

FIO Write FIO Read

0 20 40 60 80

100

0 25 50 75 100Thr

ough

put (

MB/

s)

Percentage of Duplicates

EXT4SmartDedup

DmdedupCAFTL

0 30 60 90

0 25 50 75 100Thro

ughp

ut (M

B/s)

Percentage of Duplicates

200 300

E.g., SmartDedup achieves 16.9% write speedup and 38.2% read speedup vs EXT4 with 25% of duplicates

FIO (Resource Usage)

21

Small CPU overhead (3.3%) vs. EXT4

Save battery (49.2%) vs. F2FS

Small memory footprint: 3.5MB

CPU Load Battery usage

0 200 400 600 800

1000 1200 1400

Mix 1 Mix 2Ba

ttery

Us

age(

W *

s) 0 5

10 15 20 25 30 35 40

Mix 1 Mix 2

CPU

Load

(%) EXT4

EXT4 (SmartDedup)F2FS

F2FS (SmartDedup)

I/O(GB)

Read/write ratio

Duplication ratio (%)

Mix 1 4 1 25

Mix 2 6 2 25

Trace Replay (on Nexus)

22

0%

20%

40%

60%

80%

Segment 1 Segment 2 Segment 3

Deduplication ratioWrite speedupStorage savingWrite reduction

ID Trace src

Write(GB)

Duplication ratio (%)

Read/write ratio

1 4 17.2 75.8 1.5

2 6 12.4 47.9 2.2

3 2 9.1 26.4 6.8

SmartDedup achieveso Up to 51.1% write speedup

o Up to 70.9% write reduction

Adaptive Deduplication (on Nexus)

23

Adapting to duplication level of current workload

0

200

400

600

800

0 40 80 120 160

Po

we

r C

on

sum

ptio

n (

mW

)

Time (s)

EXT4SmartDedup (basic)

SmartDedup (adaptive)

SmartDedup (adaptive)o Saves 14% power

overheado With only 8% loss

in deduplication ratio

Conclusions

Deduplication is important to smart devices—performance, utilization, endurance

SmartDedup achieves significant performance and endurance improvement with low resource cost

Trace: http://visa.lab.asu.edu/traces

24SmartDedup

Acknowledgementso Colleagues @ the ASU VISA Lab o National Science Foundation

Welcome to SmartDedup’sposter tonight 6:30pm

SmartDedup: Optimizing Deduplication for Resource ......Virtualized Infrastructures, Systems, &...

Documents

Transcript of SmartDedup: Optimizing Deduplication for Resource ......Virtualized Infrastructures, Systems, &...