Leveraging Smart Phones to Reduce Mobility Footprints Stephen Smaldone †, Benjamin Gilbert ‡, Nilton Bila *, Liviu Iftode †, Eyal de Lara *, and Mahadev.

Leveraging Smart Phones to Reduce Mobility Leveraging Smart Phones to Reduce Mobility FootprintsFootprints

Stephen Smaldone†, Benjamin Gilbert‡, Nilton Bila*, Liviu Iftode†, Eyal de Lara*, and Mahadev Satyanarayanan‡

†Rutgers University, ‡Carnegie Mellon University, *University of Toronto

June 23, 2009 MobiSys 2009 2

Traditional Mobile ComputingTraditional Mobile Computing

• Carry-Everything (on your laptop)• Primary advantage – Have all personal data, files, applications,

preferences, etc., instantaneously/immediately available • Primary disadvantage – Too “heavy”

June 23, 2009 MobiSys 2009 3

Carry-Nothing: Internet Suspend/ResumeCarry-Nothing: Internet Suspend/Resume®® (ISR) (ISR)

• Use stateless VM-enabled client stations and suspend/resume state from a server over the Internet

• ISR Parcel = VM State (memory image, disk, etc.)• Primary advantage – Carry nothing• Primary disadvantage – Depends on connectivity

ISR Client Station #1 ISR Client Station #2

SuspendResumeISR Parcel

ISR Parcel

ISR Server

June 23, 2009 MobiSys 2009 4

Existing ISR OptimizationsExisting ISR Optimizations

• On-demand disk fetching• Look-aside caching / Content-addressable storage

• Certain problems still persist– Does not work in the absence of Internet connectivity– Suspend and Resume latencies can be intolerable to users

June 23, 2009 MobiSys 2009 5

The Opportunity: SmartphoneThe Opportunity: Smartphone

• Device that people already carry• Provides ample storage• Provides multiple modes of connectivity• Light-weight (very small mobility footprint)

June 23, 2009 MobiSys 2009 6

Horatio: Mobile Self-Cleaning CacheHoratio: Mobile Self-Cleaning Cache

Fast Suspend Lazy Self-Cleaning Multiple Resume Options

3G / WiFi

Time

ISR Client Station #1 ISR Client Station #2

ISR Server

June 23, 2009 MobiSys 2009 7

Our ContributionsOur Contributions

1. Mobile self-cleaning cache (Horatio)

2. Design and prototype implementation on smartphones

3. Evaluation that demonstrates usability benefits

4. Suggestions to improve current smartphones for better mobile self-cleaning cache performance

June 23, 2009 MobiSys 2009 8

OutlineOutline

• Introduction• Horatio Design and Implementation• Evaluation and Results• Related Work• Conclusions and Future Work

June 23, 2009 MobiSys 2009 9

Horatio Design GoalsHoratio Design Goals

1. Reduce Latency– Fast suspends and resumes

2. Preserve ISR Reliability– Clean state as soon as possible

3. Conserve Battery and Storage– Store and transfer as little as possible

Key Design Principle:• Separation of Control and Data

June 23, 2009 MobiSys 2009 10

Control/Data SeparationControl/Data Separation

• Ownership Nonce – 10 bytes• Keyring – 5.5 MB• Configuration File – 500 bytes

• Ownership Nonce – 10 bytes• Keyring – 5.5 MB• Configuration File – 500 bytes

ISR Parcel

• Disk Image – 4 GB• Memory Image – 200 MB

• Disk Image – 4 GB• Memory Image – 200 MB

Parcel Control Parcel Data

• Small• Defines parcel ownership• Trusted (used to validate data)

• As large as necessary• Can be (partially) replicated• Encrypted• Possibly untrusted

June 23, 2009 MobiSys 2009 11

Control/Data Separation: How it WorksControl/Data Separation: How it Works

Fast+Lazy Suspends Efficient Self-Cleaning Efficient Resume

Parcel Data

Parcel Control

Time

ISR Server

June 23, 2009 MobiSys 2009 12

Control/Data Separation: BenefitsControl/Data Separation: Benefits

• Saves Horatio’s battery– At suspend: state transferred to Horatio (acts as receiver)

– During self-cleaning: control state transferred from Horatio (acts as transmitter)

– At resume: data state can be transferred from server

• Client can transfer state to server without impacting suspend time– Client station can transfer data state to server after user

leaves (lazy suspend)

– Trust is not an issue, Horatio will validate later (using control state)

June 23, 2009 MobiSys 2009 13

Additional Horatio OptimizationsAdditional Horatio Optimizations

• Memory image differencing– Transfer only dirty memory state as diffs during

suspends and resumes– Require basis memory image to be cached at resume

site

• Eager state transfer during client session– Transfer dirty state in the background– May result in transferring more data than necessary

due to overwrites

June 23, 2009 MobiSys 2009 14

OutlineOutline

• Introduction• Horatio Design and Implementation• Evaluation and Results• Related Work• Conclusions and Future Work

June 23, 2009 MobiSys 2009 15

Evaluation GoalsEvaluation Goals

1. How much does Horatio improve user experience?

2. How effective is self-cleaning in reducing the vulnerability of a Horatio device?

3. What is the impact of Horatio on a user’s smartphone battery?

4. How effective is eager state transfer in reducing suspend latency?

June 23, 2009 MobiSys 2009 16

Evaluation SetupEvaluation Setup

• ISR Parcel– 512 MB Memory Image and 4 GB Disk Image

• ISR Client– 2.33 GHz Core 2 Duo CPU, 3 GB RAM, 250 GB

SATA Disk, USB 2.0, Linux 2.6

• ISR Server– Dual 2.8 GHz Xeon CPUs, 1 GB RAM, 32 GB SCSI

Disk, Linux 2.6

• Horatio Devices– Nokia N95, Openmoko Freerunner, and USB

MicroSD Card

June 23, 2009 MobiSys 2009 17

Improvement in User ExperienceImprovement in User Experience

Dirty State Size (MB)

Horatio Device 1 10 100 500

ISR-1 (No Horatio) 24 25 35 82

N95-WiFi 0.7 1.2 5 21

OM-WiFi 0.6 0.8 4 17

N95-USB 0.5 0.7 2 10

SD-USB 0.4 0.4 0.7 2

Microbenchmark Suspend Results

*Values are time measured in minutes

June 23, 2009 MobiSys 2009 18

Improvement in User ExperienceImprovement in User Experience

Dirty State Size (MB)

Horatio Device 0 1 10 100 500

ISR-1 (No Horatio) 5 - - - -

N95-WiFi - 7 7 9 16

OM-WiFi - 5 5 6 12

N95-USB - 4 4 5 9

SD-USB - 0.6 0.6 0.6 0.9

Microbenchmark Resume Results

*Values are time measured in minutes

June 23, 2009 MobiSys 2009 19

Realistic Workloads (Macrobenchmarks)Realistic Workloads (Macrobenchmarks)

Workload Name Execution Time (min)

Dirty State (MB)

Memory Disk

Email 6 16 3

Word 10 41 3

Photo 13 25 4

Shop 12 31 14

Podcast 7 120 109

Video 40 264 368

June 23, 2009 MobiSys 2009 20

Improvement in User ExperienceImprovement in User Experience

0

500

1000

1500

2000

2500

3000

ISR-1 ISR-10 N95-WiFi SD-USB

Email

Tim

e (

s)

suspend

execution

resume

Macrobenchmark Results

0

2000

4000

6000

8000

10000

12000

ISR-1 ISR-10 N95-WiFi SD-USB

Video

Tim

e (

s)

suspend

execution

resume

June 23, 2009 MobiSys 2009 21

Self-Cleaning TimeSelf-Cleaning Time

Dirty State Size (MB)

Horatio Device 1 10 100

N95-WiFi 0.6 minutes 2 minutes 15 minutes

OM-WiFi 0.2 minutes 1 minute 13 minutes

N95-3G 3 minutes 8 minutes 1 hour

Workload (Dirty State Size) N95-WiFi N95-3G

Email (20 MB) 4 minutes 12 minutes

Word (44 MB) 16 minutes 1 hour

Photo (29 MB) 14 minutes 1 hour

Shop (44 MB) 18 minutes 1 hour

Podcast (230 MB) 37 minutes 2 hours

Video (632 MB) 2 hours 7 hours

June 23, 2009 MobiSys 2009 22

Horatio Battery ConsumptionHoratio Battery Consumption

Dirty State Size (MB)

Operation Horatio Device 1 10 100 500

Suspend N95-WiFi 0.2% 0.4% 3% 11%

Suspend N95-USB 0.1% 0.2% 0.9% 4%

Resume N95-WiFi 3% 4% 5% 9%

Resume N95-USB 0.6% 0.6% 0.7% 1%

Self-Clean N95-WiFi 0.2% 0.6% 6% -

Self-Clean N95-3G 1% 4% 28% -

*Values are percentage of battery depleted

June 23, 2009 MobiSys 2009 23

Eager State PropagationEager State Propagation

Workload Lazy State

(MB)

Suspend State

(MB)

Eager State

(MB)

Email 20 3 130

Word 44 2 221

Photo 28 2 199

Shop 44 29 486

June 23, 2009 MobiSys 2009 24

Related WorkRelated Work

Remote(low physical vulnerability)

Storage Site

Local(high network resilience)

Lo

cal

(cris

p in

tera

ctio

n)R

emo

te(h

igh

com

pute

pow

er)

Ex

ecu

tio

n S

ite

Classic PC modelLaptops

SoulPad, MojoPac

Remote ExecutionCyber Foraging

ISR model

Thin client model

Snowbird Transient Thin Client

Ho

rati

o

June 23, 2009 MobiSys 2009 25

Conclusions and Future WorkConclusions and Future Work

• Reliable and Efficient Carry-Nothing Mobile Computing– Use the smartphone as a self-cleaning mobile cache

to improve ISR

• Experimental Results Demonstrate:– Suspend and resume latencies reduced up to 98%

• Future Work– Prediction of resume location for state prefetching– Horatio user interface– Deployment

Thank You!Thank You!

June 23, 2009 MobiSys 2009 27

Impact on Mobility FootprintImpact on Mobility Footprint

Dirty State Size (MB)

Operation Horatio Device 1 10 100 500

Suspend N95-WiFi 28 71 400 1789

Suspend N95-USB 12 31 147 609

Resume N95-WiFi 507 613 756 1456

Resume N95-USB 96 97 120 227

Self-Clean N95-WiFi 36 103 916 -

Self-Clean N95-3G 181 565 4553 -

*Values are energy measured in Joules

June 23, 2009 MobiSys 2009 28

Workload State GenerationWorkload State Generation

0

10

20

30

40

50

60

70

1 2 3 4 5 6 7 8 9 10 11 S

Cleaning Cycles

Dir

ty S

tate

(M

B)

Email

Word

Photo

Shop

June 23, 2009 MobiSys 2009 29

Update LocalityUpdate Locality

0

20

40

60

80

100

0 2 4 6 8 10 12 14

Number of Updates

% o

f M

em

ory

Lo

ca

tio

n U

pd

ate

s

Email

Word

Photo

Shop