weardrive: fast and energy efficient storage for wearables

WearDrive: Fast and Energy Efficient Storage for Wearables

Reza Shisheie

Cleveland State University

CIS 601

2

Wearable Computing: A New Era

……

2

Wearable Computing: A New Era

……

Fitness/HealthcareNotifications Audio/Video Data Analytics

……

3

Wearables Are Battery Constrained

+1.5 - 2.5K mAh

3


+1.5 - 2.5K mAh

f (Size & Weight)

+100 – 500 mAh

3


+1.5 - 2.5K mAh

f (Size & Weight)

+100 – 500 mAh

Mobile SoC,

OS & SDK

Wearable SoC,

OS & WearSDK

Reduce

Development Cost

3


+1.5 - 2.5K mAh

f (Size & Weight)

+100 – 500 mAh

Mobile SoC,

OS & SDK

Wearable SoC,

OS & WearSDK

Reduce

Development Cost

Offload Wearable’s Workload to the Phone

4

Energy Challenge: Wearable Apps Are Data-Intensive

Extended Display

……

Notifications

Sensors

……

16+ Sensors

4


Extended Display

……

Notifications

Sensors

……

16+ Sensors

4


Extended Display

……

Notifications

Fast & Energy-Efficient

Storage System

5

Mobile Storage is Energy-Intensive

Flash device is not energy-intensive,

Storage software stack is energy-intensive !

5


0

20

40

60

80

100

120

Random Read Random Write Sequential Read Sequential Write

Energ

y O

verh

ead

Facto

rs

4 KB 32 KB

Storage software consumes 80–110x more energy

than Flash [Li et al., FAST’14]



5




Slow Flash:

Increased CPU Idle Time

+

Runtime System

Overhead

+

Data

Encryption

6

Use DRAM as Storage?

• DRAM is fast

• Closer to applications

• No software overhead for isolation and security

OS

Flash

DRAM

Applications

6

Use DRAM as Storage?

• DRAM is fast

• Closer to applications

• No software overhead for isolation and security

OS

Flash

DRAM is volatile, the data durability is a problem !

DRAM

Applications

7

Leveraging Batteries for Non-Volatility

DRAM

Battery

= Non-Volatile

Memory

7


DRAM

Battery

=BB-RAM

Non-Volatile

Memory

7


DRAM

Battery

=

No Hardware

Changes

BB-RAM

Non-Volatile

Memory

7


DRAM

Battery

=

No Hardware

Changes

+Performance

Improvement

BB-RAM

Non-Volatile

Memory

7


DRAM

Battery

=

No Hardware

Changes

+Data

Durability

+Performance

Improvement

BB-RAM

Non-Volatile

Memory

8

Our BB-RAM Implementation

BB-RAM page DRAM page

8



BB-RAM Manager

8



BB-RAM Manager

Memory

PressureAutomatically adjust BB-RAM size

according to application requirement

8



BB-RAM Manager

OS

Crashes

8



BB-RAM Manager

OS

DeadlockA watchdog timer to

detect if OS is hung

9

Leveraging Phone for Capacity/Compute

Wearable applications focus on the latest data,

but how about the old data?

9

Leveraging Phone for Capacity/Compute

Writing data to remote memory via WFD is more energy efficient

0

0.1

0.2

0.3

0.4

4 8 16 32

mJ/K

B

IO Size (KB)

Local Flash Remote Memory via WiFi Direct

Two Moto E phones with 802.11 b/g/n

Wearable applications focus on the latest data,

but how about the old data?

10

Reaching the Phone Efficiently

Standby Power Data Transfer

BLE Low 0.41 microjoules/bit

WFD High 0.02 microjoules/bit

BLE for small data transfer, WFD for large data transfer

BLE

WiFi Direct

10

Reaching the Phone Efficiently

Standby Power Data Transfer

BLE Low 0.41 microjoules/bit

WFD High 0.02 microjoules/bit

Use BLE’s connection quality for reference

BLE for small data transfer, WFD for large data transfer

BLE

WiFi Direct

11

Building A Distributed In-Memory Storage System

Application

Wearable Phone

FlashEfficient

Wireless

Network

2 Reduce Flash size reduce $ cost (10%)

Improve performance & save energy1

3 Leverage phone’s resources for wearable

Cache

BB-RAM

BB-RAM

Storage

+

12

WearDrive Design: Key Value Store

Application

Wearable Phone

FlashEfficient

Wireless

Network

WearCache

BB-RAM

BB-RAM

WearKV

+

12

WearDrive Design: Key Value Store

Application

Wearable Phone

FlashEfficient

Wireless

Network

WearCache

BB-RAM

BB-RAM

WearKV

+

Why Key Value

Store?

Focus on

Latest Data

F I F O

Hybrid

Networks

+ +

13

Putting It All Together

Hybrid

BLE/WFD

WearCache WearKV

Wearable Phone

13


Hybrid

BLE/WFD

WearCache

<key, value>

WearKV

Wearable Phone

13


Hybrid

BLE/WFD

WearCache

<key, value>

WearKVNotification

Wearable Phone

13


Hybrid

BLE/WFD

<key, value>

WearCache

<key, value>

WearKVNotification

Wearable Phone

13


Hybrid

BLE/WFD

WearCache WearKV

Wearable Phone

13


Hybrid

BLE/WFD

WearCache WearKV

RegisterForSensor (Sensor ID)

Wearable Phone

13


Hybrid

BLE/WFD

<key, value>

WearCache WearKV


Wearable Phone

13


Hybrid

BLE/WFD

<key, value>

WearCache

<key, value>

WearKV


Wearable Phone

13


Hybrid

BLE/WFD

<key, value>

WearCache

<key, value>

WearKV


RegisterCallBack (TimeGap,

CallBackFun)

Wearable Phone

13


Hybrid

BLE/WFD

<key, value>

WearCache

<key, value>

WearKVNotification


RegisterCallBack (TimeGap,

CallBackFun)

Wearable Phone

14

WearDrive: Implementation in Real System

Mobile Device

Android OS (version 4.4)

Application

14


Mobile Device


KV-Store:

WearCache/WearKV

BB-RAM

Manager

Hybrid Network

Management

Application

14


Mobile Device


KV-Store:

WearCache/WearKV

BB-RAM

Manager

Hybrid Network

Management

WearDrive API

Application

15

Experimental Setup

Emulated wearable device (similar to Samsung Gear)

1.2 GHz Dual-core + 512 MB RAM + BLE 4.0 + WiFi 802.11 b/g/n

Monsoon power monitor

15

Experimental Setup

WearableOnly Use local Flash as storage

WearSDK★ Android Wear SDK

WearDrive In-memory storage system

★Its data layer was expanded to make it support WFD and hybrid network protocol.

16

WearBench: Benchmarks for Wearables

Examples Parameters

Extended

DisplaySize, inter-arrival time

SensorsSampling rate,

monitoring period

Audio/VideoEncoding rate, quality,

monitoring period

17

Local Memory vs. Local Flash (Write)

0

5

10

15

128 256 512 1024

Thr

ough

put (

K T

PS

)

IO Size (Bytes)

WearableOnly WearDrive

1 2 4

1

2

4 threads

0

20

40

60

80

128 256 512 1024

Ene

rgy

(Jou

les)

9x performance improvement, 4x less power

18

Local Memory vs. Local Flash (Read)

2x performance improvement, 1.7x less power

0

5

10

15

20

25

128 256 512 1024

Thr

ough

put (

K T

PS

)

IO Size (Bytes)

WearableOnly WearDrive

12

4 1

2

4 threads

0

1

2

3

4

5

128 256 512 1024

Ene

rgy

(Jou

les)

19

Energy Usage of Sensor Data Aggregation

0

500

1000

1500

2000

2500

3000

WearableOnly WearSDK-BLE WearSDK-WFD WearSDK-HYN WearDrive

Ene

rgy

(Jou

les)

Storage Networking

WearDrive consumes 1.5x less power than WearableOnly,

up to 3x less power than WearSDK

The sensor samples values continuously at 1 Hz for 24 hours

20

Energy Usage of Receiving Notifications

0

2

4

6

8

10

12

5 10 15 20 30 40 50 60

Ene

rgy

(Jou

les)

Interval between two notifications (seconds)

WearSDK-BLE WearSDK-WFD WearSDK-HYN WearDrive

WearDrive saves 1.2 - 3x energy than WearSDK-BLE

21

Compute Offload & Impact on the Phone

Wearable (300 mAh) Phone (2000 mAh)

WearableOnly 27.12% --

WearableOnly

+InMemory13.23% --

WearDrive 0.87% 2.09%

Data analysis with k-means on the data set generated by

16 sensors at 1 Hz for 24 hours.

22

Conclusion

2

In-memory Storage System

with battery-back DRAM (BB-RAM)

Extended Capabilities

leveraging phone’s storage and CPU

1

3Extended Lifetime

3x improvement on battery life time

WearDrive: Storage System for Wearables

†

Thanks!

Q&A

weardrive: fast and energy efficient storage for wearables

Documents