Ghosts in the Machine:Interfaces for Better Power Management

Manish Anand

Edmund B. Nightingale

Jason Flinn

Electrical Engineering and Computer Science Dept.University of Michigan

Manish Anand

2

Need for improved power management

Capabilities of mobile handheld devices improving rapidly: Wireless connectivity Storage capacity

Battery capacity improving slowly I/O devices decrease handheld battery lifetime by 60% OEMs provide techniques to reduce battery consumption by

severely compromising the performance Careful power management needed

Manish Anand

3

Tradeoff between energy and performance

Inherent tradeoff exists between performance and energy conservation

Goals of power management involve:

Lower energy usage

Lower interactive response time (performance)

Time

En

erg

y

Good performance

Extended battery life

Manish Anand

4

Current device power management

better performance and energy conservation

simplicityMODULARITY

moreless

Problem: Device-centric strategies do not consider the operating environment contexts, such as Base power of mobile computer Activity of other devices Application intent

Our Approach: Provide an infrastructure to expose additional information about operating environment to both applications and devices

Manish Anand

5

Goals and contributions

Goals: Improve both performance and energy Enable cross-device optimization Allow end users to specify their relative priorities for

performance and energy conservation

Contributions: Simple interfaces to expose additional contexts Power manager for querying device state Self-tuning power management (STPM) modules for disk, n/w Ghost hints for cross-device power management

Manish Anand

6

Outline Motivation and background Overview of our solution Interfaces for better power management Adaptive caching Evaluation Related work and conclusions

Manish Anand

7

Review: power management Network Device: 802.11b wireless

CAM – Continuously Active Mode PSM – Power Saving Mode

+ power usage reduced by 70 to 80%

- delay proportional to polling period (100 ms)

Disk Device: IBM/Hitachi microdrive Active Low Power Idle / Standby

+ power usage reduced by 55% / 90%

- transition time of 300ms / 800ms

Manish Anand

8

Impact of power management

Assumption: Fetching data from local storage is less costly

Web Fetch Time

0

500

1000

1500

2000

2500

0 100 200 300 400 500 600

File Size (KB)

Fet

ch t

ime

(ms)

Disk Standby

Disk Active

Network PSM

Network CAM

Local data access could be more expensive for small files Break-even point is dependent on the state of the device Simple adaptation through cache hierarchy

Manish Anand

9

Limitation of adaptation

Example: Browser fetching many small images For single image fetching from network is correct For many images fetching from disk is correct

Amortized transition cost over many reads

Problem: Reactive adaptive caching does not help Disk continues to be in standby state

Solution: Apps disclose “accesses that might have been” We use ghost hints to implement this

Manish Anand

10

Outline Motivation and background Overview of our solution Interfaces for better power management Adaptive caching Evaluation Related work and conclusions

Manish Anand

11

Energy-aware architecture

Applications

Power Manager

NetworkHints

DeviceCharacteristics & Current State

STPMNetwork

STPMDisk

ModeTransitions

ModeTransitions

Cache Manager

Ghost HintsGhost Hints

Data AccessDecision

Notify NetworkState Change

Notify DiskState Change

Operating System

Disk Device Driver

Network Device Driver

MobiCom ‘03

Manish Anand

12

Power manager Central repository that maintains information about:

State of the device Performance characteristics of I/O devices Energy characteristics of I/O devices

Power manager interface 3 calls for devices – register, deregister and notify 5 calls for applications – 4 to query device state and

characteristics, 1 for registering callbacks

Manish Anand

13

Self-tuning power management STPM modules consider:

Application access patterns Base power of mobile computer Energy and performance characteristics of the device Relative priority of performance and energy conservation

STPM modules decide: Time to transition to high power state Time to transition to low power state

Energy conservation

Performance0 100knob

Manish Anand

14

Disk power management with STPM

Transition to power saving mode employs Break-even heuristic

Disk is likely to remain idle if it has been idle for a period Incorporate the STPM principles

Transition to active mode employs both Reactive strategy

Transition on a request Proactive strategy

Use ghost hints to transition even with no device access

Manish Anand

15

Ghost hints

Ghost Hint Ghost Hint

Expended Energy between hints

Break-even Threshold

Ghost hints contain opportunity cost in terms of time and energy Issued by the application, when a poor I/O path is chosen due to inappropriate device state

Decrements energy cost of being active between hints Transitions to high performance state on an overflow

state change on overflow

Manish Anand

16

Cache manager

Keeps the meta-data information in-memory hash table Determines the weighted cost of going to network or disk Maintains a write queue and delays writes to the disk

ApplicationsCache

Manager

Disk

CacheStatus?

network bettern/w fetch

data PutData

flush

ghost hint

STPMModule

PowerManagerdevice state

Manish Anand

17

Outline Motivation and background Overview of our solution Interfaces for better power management Adaptive caching Evaluation Related work and conclusions

Manish Anand

18

Evaluation Goals:

Effect of cache manager’s adaptive, cross-device strategy

Influence of global knob

Benefits of ghost hints

Client: iPAQ handheld with Cisco 350 wireless card and IBM/Hitachi microdrive

Test applications:

Email-sync - recorded e-mail activity

Dillo web browser with wwwoffle - Berkeley web traces

Manish Anand

19

Effectiveness of energy-aware caching

0

100

200

300

400

500

600

700

Average Response Time

(ms)

E-mail Web

Static

Adaptive

0

500

1000

1500

2000

2500

3000

Energy (Joules)

E-mail Web

Equal priority for performance and energy Average response time is reduced by 27% to 42% Total energy usage is reduced by 5% to 9%

Manish Anand

20

E-mail and web with no ghost hints

STPM modules base their decision solely on device accesses No change for knob value less than 95

k=100

k=99

k=95-0 static

2000

2500

3000

3500

4000

4500

5000

0 200 400 600

Average Response Time(ms)

En

erg

y(J

)

k=100

k=99

k=95-0static

120014001600

1800200022002400

260028003000

0 200 400 600

Average Response Time(ms)

En

erg

y(J

)

E-mail Web

Manish Anand

21

Importance of ghost hints

Email: Substantial benefit when performance is high priority Web: Less likely to see run of accesses clustered together

k=100

k=99

k=95k=90-0

k=99

k=95-0 static

2000

2500

3000

3500

4000

4500

5000

0 200 400 600

Average Response Time(ms)

En

erg

y(J

)

k=100

k=99

k=95

k=70-0k=90

k=99

k=95-0static

1200140016001800200022002400260028003000

0 200 400 600

Average Response Time(ms)

En

erg

y(J

)

E-mail Web

Manish Anand

22

Importance of ghost hints

static

knob=50-0knob=70knob=90

knob=95

knob=99

knob=95-0

knob=99

knob=100

2000

2500

3000

3500

4000

4500

5000

0 200 400 600

Average Response Time(ms)

En

erg

y(J)

Web with full cache

Ghost hints show a positive effect on the system

Ghost hints yield substantial benefit for some workloads, and do no harm in the situations where they seem ineffective

Manish Anand

23

Related work We are the first system to efficiently access data on multiple I/O

Burstiness for energy efficiency (Weissel ’02, Papathanasiou ’03, Heath ‘02)

ACPI: similar to power manager but too complex, does not disclose device characteristics

Wake on Wireless (Shih ’02)

OS level power management (ECOSystem, Odyssey)

Manish Anand

24

Conclusions

Contributions: Improved power management interfaces: Power manager for querying device state STPM modules for disk, network Ghost hints for cross-device power management

Current work: Modification of cached data Multiple network and storage devices Feedback loop to dynamically set the global knob

Ghosts in the Machine:Interfaces for Better Power Management

Manish Anand

Edmund B. Nightingale

Jason Flinn

Electrical Engineering and Computer Science Dept.University of Michigan

Manish Anand

26

Dynamic Power Management

The energy used by the I/O devices can be prohibitive, without power managementHitachi Microdrive Cisco 350 wireless card

Mode Power Mode Power

Performance Idle 760 mW CAM 1410 mW

Low-Power Idle 340 mW PSM 390 mW

Standby 80 mW

Battery lifetime is decreased by 60% without power management

Drawback of using power management Network device

Delay proportional to the polling period Disk device

Transition cost to switch to active mode (300ms to 800ms)

ghostrealized disk spin-up

PSMswitch

CAMswitch

Network access Disk access

ghost hints STPM disk hints

STPM network hints

Applications

Disk

Network

Manish Anand

28

Network Power Management with STPM

STPM network switches from PSM to CAM when: Application specifies max delay < beacon period Disclosed transfer size > break-even size Many forthcoming transfers are likely

To predict forthcoming transfers STPM network generates an empirical distribution of run lengths

>150 ms >150 ms>150 ms

Transfers

Run Run Run Run