APE:AnAnnota*onLanguageand% …ochipara/presentations/ICSE2014.pdf · NetAct" OR 3G 4G. TheAPEPolicyModel ... 1 3 7 15 31 63 127 191 255 383 511 % StateExpressionLength Overhead"of"IPC"measured"to"be"1.71"ms"

APE: An Annota*on Language and Middleware for Energy-‐Efficient Mobile Applica*on Development

Nima Nikzad, Octav Chipara†, and William G. Griswold

[email protected] Dept. of Computer Science and Engineering, UC San Diego

Dept. of Computer Science, University of Iowa†

Growth of Con*nuously-‐Running Mobile Applica*ons

• Runs in background, out of user’s sight • Periodic workloads • OIen delay-‐tolerant •  e.g., health monitoring, sports and news feeds, social networking

2

How can we run many CRM applica3ons while minimizing impact on ba:ery life?

Low-‐ and Applica*on-‐level Op*miza*ons are Both Needed

•  Low-‐level Op2miza2ons can only do so much

•  DVFS, Nckless kernels, radio scheduling, batched writes to flash, etc…

• Applica2on-‐level Op2miza2ons are necessary • Workload shaping, filtering, piggy-‐backing radio transmission, etc… •  Take advantage of applica2on specific behavior

3

Thread uploadThread = new Thread(new Runnable() { while(true) { try { Thread.sleep(1200000); // Sleep 20 min. } catch(InterruptedException e) { attemptUpload(); } }); uploadThread.start();

Case Study: Naïve Upload in Ci*Sense

4

CO, NO2, O3

Alerts, Maps

Small Transmissions à Big Impact

5

t

Power

6 sec.

12 sec.

1 W

10 mW

t

Power

Piggybacking is Effec*ve for Amor*zing Cost of Communica*on

6

1 W

10 mW

Thread uploadThread = new Thread(new Runnable() { while(true) { Intent batt = context.registerReceiver(null, new IntentFilter(Intent.ACTION_BATTERY_CHANGED)); int lvl = batt.getIntExtra(BatteryManager.EXTRA_LEVEL, -‐1); int scl = batt.getIntExtra(BatteryManager.EXTRA_SCALE, -‐1); float batteryPct = lvl / (float) scl; try { if(batteryPct > 70) { Thread.sleep(1200000); } else { Thread.sleep(3600000); } } catch(InterruptedException e) { attemptUpload(); } }); uploadThread.start(); TelephonyManager teleManager = (TelephonyManager) context.getSystemService(Context.TELEPHONY_SERVICE); TransListener transListener = new TransListener(); teleManager.listen(transListener, PhoneStateListener.LISTEN_DATA_ACTIVITY); private class TransListener extends PhoneStateListener { public void onDataActivity(int act) { if(act == TelephonyManager.DATA_ACTIVITY_IN || act == TelephonyManager.DATA_ACTIVITY_OUT || act == TelephonyManager.DATA_ACTIVITY_INOUT) { uploadThread.interrupt(); } } }

Code is Quickly Complicated by Power-‐Management Policy

7

Thread uploadThread = new Thread(new Runnable() { while(true) { try { Thread.sleep(1200000); } catch(InterruptedException e) { attemptUpload(); } }); uploadThread.start();

Thread uploadThread = new Thread(new Runnable() { while(true) { try { Thread.sleep(1200000); } catch(InterruptedException e) { attemptUpload(); } }); uploadThread.start(); TelephonyManager teleManager = (TelephonyManager) context.getSystemService(Context.TELEPHONY_SERVICE); TransListener transListener = new TransListener(); teleManager.listen(transListener, PhoneStateListener.LISTEN_DATA_ACTIVITY); private class TransListener extends PhoneStateListener { public void onDataActivity(int act) { if(act == TelephonyManager.DATA_ACTIVITY_IN || act == TelephonyManager.DATA_ACTIVITY_OUT || act == TelephonyManager.DATA_ACTIVITY_INOUT) { uploadThread.interrupt(); } } }

Building Energy-‐Efficient Apps is Not Simple!

• Code for power-‐management tends to be complex •  Highly applicaNon specific

•  Trade-‐off: Energy vs Timeliness and/or Accuracy •  Thread management, event and state change handling

• OpNmizaNons should be postponed unNl applicaNon requirements are set •  Use cases and ‘acceptable’ delay may change over Nme •  But… Retrofiengà Time + Bugs

8

What Do Energy-‐Management Policies Typically Look like?

• We examined the latest research, best-‐pracNce guides, documentaNon, and open-‐source projects •  Android Best PracNce Guide, AT&T Research, [Wang09], [Nath12], [Nikzad12], [Musolesi10], Cyanogen Project

• Common thread: Timing and Device State!

•  “Wait unNl _______ before execuNng _______”

9

APE is a Declara*ve Annota*on Language and Run*me for Policies

10


Thread uploadThread = new Thread(new Runnable() { while(true) { @APE_If(“Battery.Level > 70%”) @APE_WaitUntil(“Network.Active”, MaxDelay=1200) @APE_Else() @APE_WaitUntil(“Network.Active”, MaxDelay=3600) attemptUpload(); } }); uploadThread.start();

Rest of this Talk

11

cG ≥ 20 min : true

start

true : Net.Activity

while(true) { @APE_WaitUntil(“Net.Active”, MaxDelay=1200) uploadSensorData(); }

OR

WiFi AND

OR NetAct

3G 4G

The APE Policy Model EffecNvely Modeling Policies Using Timed Automata •  Clarify semanNcs

•  Drive design of language and runNme

12

Many Policies Can Be Expressed Using Simple Timed Automata

13

Wait up to 20 minutes for cellular network ac5vity


start

true : Net.Activity

P = ({(Accel.Move, ∞), (GPS.Drive, 30 sec)}, ∞)

More Complicated Policies Can Fall Back to Earlier States

Wait for movement using the accelerometer, then wait up to thirty seconds for driving to be detected

14

true : Accel.Move

cAcc ≥ 30 sec : true

true : GPS.Drive start

Many Policies Can Be Expressed Using Very Simple Timed Automata


15

P = ({(Net.Active, ∞)}, 20 min)


start

true : Net.Activity

The APE Annota*on Language RepresenNng Power-‐management Policies as AnnotaNons •  Expressing simple Nmed automata using text

16

while(true) { @APE_WaitUntil(“Net.Active”, MaxDelay=1200 PreWait=“log(‘Started waiting for activity…’)”, PostWait=“log(‘Activity detected!’)”) uploadSensorData(); }

while(true) { @APE_WaitUntil(“Net.Active”, MaxDelay=1200) uploadSensorData(); }

APE_WaitUn*l: Wait For Desired Device State

17

P = ({(Net.Active, ∞)}, 20 min)


APE_WaitUn*l: Wait For Desired Device State

18

P = ({(Accel.Move, ∞), (GPS.Drive, 30 sec)}, ∞)

while(true) { @APE_WaitUntil(“({accMove()},inf), ({gpsDrive()},30)”, MaxDelay=inf) downloadDriveData(); }

Wait for movement using the accelerometer, then wait up to thirty seconds for driving to be detected

Other APE Constructs •  if-‐else constructs

• CreaNng new terms and saving policies

19

while(true) { @APE_If(“Battery.Level > 70%”) @APE_WaitUntil(“Net.Active”, MaxDelay=1800) @APE_Else() @APE_WaitUntil(“WiFi.Connected”, MaxDelay=3600) uploadSensorData(); }

@APE_DefineTerm(“MyTerm”, “Battery.Charging AND (WiFi.Connected OR Cell.4G)”)

@APE_DefinePolicy(“MyPolicy”, “(Display.Off,inf),(MyTerm,10)”)

The APE Middleware Service RunNme Hardware Monitoring on Behalf of Client Apps •  Minimizing overhead associated with monitoring

20

Annota*ons are Processed at Compile-‐*me into Run*me Requests

21

APE Annotated Source

Source Including Generated

APE Requests

APE AnnotaNon Preprocessor

Clients Start Up the APE Service and Send Requests at Run*me

22

Client Process A

Android System

Bind me to the APE Service APE Service

Client Process B

Bind me to the APE Service

WaitUnNl(1)

A.1: Net.AcNve

B.1: …

A.1: Net.AcNve

Register “Net.AcNve” as 1

APE Service

A.1: Net.AcNve

B.1: …

APE Returns Control to the Calling Thread Once the Request is Sa*sfied

23

Android System

Client Process A

Client Process B

Tell me about changes in cellular state

Data outgoing on cellular radio

Return WaitUnNl(1)

Boolean Expressions Are Transformed into Expression Trees

24

WiFi.Connected OR Network.AcNve AND (Cell.3G OR Cell.4G) OR

WiFi AND

OR NetAct

3G 4G

APE Service

The APE Service Performs Device State Monitoring on Behalf of Clients

25

OR

WiFi AND

OR NetAct

3G 4G

Network State Monitor

Android System

F

F F

F F

F F

F F T F

T

T

T

T

T

T

Evalua*on A Case Study of OpNmizing a CRM ApplicaNon Using APE •  How complicated are policies in pracNce?

•  What are the potenNal benefits and overhead?

26


Thread uploadThread = new Thread(new Runnable() { while(true) { @APE_If(“Battery.Level > 70%”) @APE_WaitUntil(“Network.Active”, MaxDelay=1200) @APE_Else() @APE_WaitUntil(“Network.Active”, MaxDelay=3600) attemptUpload(); } }); uploadThread.start();

Case Study: Ci*Sense

27

Thread uploadThread = new Thread(new Runnable() { while(true) { @APE_If(“Battery.Level > 70%”) @APE_WaitUntil(“WiFi.Connected OR Network.Active AND (Cell.3G OR Cell.4G)”, MaxDelay=1200) @APE_Else() @APE_WaitUntil(“WiFi.Connected OR Network.Active AND (Cell.3G OR Cell.4G)”, MaxDelay=3600) attemptUpload(); } }; uploadThread.start();

Four APE annota2ons versus 45 lines of Java (thread suspension/waking, new class for handling callbacks, registering for callbacks)

Device Power ConsumpNon

Background Apps Can Repeatedly Wake Resources

28

Six applicaNon instances with no piggybacking

The Simplest of Policies Can Have a Big Impact on Power-‐Consump*on

29

Six applicaNon instances with @APE_WaitUntil(“Net.Active”, MaxDelay=120)

Device Power ConsumpNon

0

100

200

300

400

500

600

0 1 2 3 4 5

Increase in Pow

er Con

sump2

on (m

W)

Number of Addi2onal Applica2ons

APE Apps Non-‐APE Apps

Increase in power consumpNon aIer first instance

The Simplest of Policies Can Have a Big Impact on Power-‐Consump*on

30

Increase of 13.49 mw (1.6%)

Increase of 551.06 mw (63.7%)

Overhead of Execu*ng Each WaitUn*l Call vs Hand-‐coded

31

0

5

10

15

20

25

1 3 7 15 31 63 127 191 255 383 511

Time to Resolve (m

s)

State Expression Length

Overhead of IPC measured to be 1.71 ms

Longest Expression from CiNSense

Limita*ons and Future Work

• Assumes developer is already aware of opportuniNes for energy savings •  Under Development: AutomaNcally idenNfying costly operaNons and suggesNng safe, effecNve policies

• User study to properly evaluate APE’s ease-‐of-‐use and sufficient expressibility of language

32

Conclusion •  Use of 2ming is widely applicable for power-‐management in mobile applicaNons

•  A simple annota2on language and run2me can make this technique conveniently accessible to programmers

•  Timed automata provide a seman2c model that informs the design of language and runNme

•  APE significantly eases the implementa2on of policies in CiNSense and achieves drama2c power-‐savings

33

Acknowledgement •  This work was supported by the NaNonal Science FoundaNon and the Roy J. Carver Charitable Trust

34

William Griswold UC San Diego

Octav Chipara Univ. of Iowa

Conclusion •  Use of 2ming is widely applicable for power-‐management in mobile applicaNons

•  A simple annota2on language and run2me can make this technique conveniently accessible to programmers

•  Timed automata provide a seman2c model that informs the design of language and runNme

•  APE significantly eases the implementa2on of policies in CiNSense and achieves drama2c power-‐savings

35

DON’T USE -‐ Anima*on back up

36

t

Power

DON’T USE -‐ Anima*on back up

37

t

Power

38

t

Power

Challenge: Managing Energy-‐Consump*on in CRM Apps

39

•  Timing plays a significant role in energy-‐consumpNon

• Waking the CPU •  Transmieng data •  Powering the display

Limited ba:ery!

t

Power

Transmit as much as possible at this Nme!

CO, N

O2, O3

Alerts, Maps

Thread uploadThread = new Thread(new Runnable() { while(true) { try { Thread.sleep(120000); } catch(InterruptedException e) { attemptUpload(); } }; uploadThread.start();


40

Thread uploadThread = new Thread(new Runnable() { while(true) { @APE_If(“Battery.Level > 70%”) @APE_WaitUntil(“Network.Active”, MaxDelay=1200) @APE_Else() @APE_WaitUntil(“Network.Active”, MaxDelay=3600) attemptUpload(); } }; uploadThread.start();

Thread uploadThread = new Thread(new Runnable() { while(true) { Intent batt = context.registerReceiver(null, new IntentFilter(Intent.ACTION_BATTERY_CHANGED)); int lvl = batt.getIntExtra(BatteryManager.EXTRA_LEVEL, -‐1); int scl = batt.getIntExtra(BatteryManager.EXTRA_SCALE, -‐1); float batteryPct = lvl / (float) scl; try { if(batteryPct > 70) { Thread.sleep(120000); } else { Thread.sleep(360000); } } catch(InterruptedException e) { attemptUpload(); } }; uploadThread.start();

Thread uploadThread = new Thread(new Runnable() { while(true) { Intent batt = context.registerReceiver(null, new IntentFilter(Intent.ACTION_BATTERY_CHANGED)); int lvl = batt.getIntExtra(BatteryManager.EXTRA_LEVEL, -‐1); int scl = batt.getIntExtra(BatteryManager.EXTRA_SCALE, -‐1); float batteryPct = lvl / (float) scl; try { if(batteryPct > 70) { Thread.sleep(120000); } else { Thread.sleep(360000); } } catch(InterruptedException e) { attemptUpload(); } }; uploadThread.start(); TelephonyManager teleManager = (TelephonyManager) context.getSystemService(Context.TELEPHONY_SERVICE); TransListener transListener = new TransListener(); teleManager.listen(transListener, PhoneStateListener.LISTEN_DATA_ACTIVITY); private class TransListener extends PhoneStateListener { public void onDataActivity(int act) { if(act == TelephonyManager.DATA_ACTIVITY_IN || act == TelephonyManager.DATA_ACTIVITY_OUT || act == TelephonyManager.DATA_ACTIVITY_INOUT) { uploadThread.interrupt(); } } }

Thread uploadThread = new Thread(new Runnable() { while(true) { Intent batt = context.registerReceiver(null, new IntentFilter(Intent.ACTION_BATTERY_CHANGED)); int lvl = batt.getIntExtra(BatteryManager.EXTRA_LEVEL, -‐1); int scl = batt.getIntExtra(BatteryManager.EXTRA_SCALE, -‐1); float batteryPct = lvl / (float) scl; try { if(batteryPct > 70) { Thread.sleep(120000); } else { Thread.sleep(360000); } } catch(InterruptedException e) { attemptUpload(); } }; uploadThread.start();

Thread uploadThread = new Thread(new Runnable() { while(true) { Intent batt = context.registerReceiver(null, new IntentFilter(Intent.ACTION_BATTERY_CHANGED)); int lvl = batt.getIntExtra(BatteryManager.EXTRA_LEVEL, -‐1); int scl = batt.getIntExtra(BatteryManager.EXTRA_SCALE, -‐1); float batteryPct = lvl / (float) scl; try { if(batteryPct > 70) { Thread.sleep(120000); } else { Thread.sleep(360000); } } catch(InterruptedException e) { attemptUpload(); } }; uploadThread.start(); TelephonyManager teleManager = (TelephonyManager) context.getSystemService(Context.TELEPHONY_SERVICE); TransListener transListener = new TransListener(); teleManager.listen(transListener, PhoneStateListener.LISTEN_DATA_ACTIVITY); private class TransListener extends PhoneStateListener { public void onDataActivity(int act) { if(act == TelephonyManager.DATA_ACTIVITY_IN || act == TelephonyManager.DATA_ACTIVITY_OUT || act == TelephonyManager.DATA_ACTIVITY_INOUT) { uploadThread.interrupt(); } } }

CO, N

O2, O3

Alerts, Maps

Thread uploadThread = new Thread(new Runnable() { while(true) { try { Thread.sleep(120000); } catch(InterruptedException e) { attemptUpload(); } }; uploadThread.start();


41

Thread uploadThread = new Thread(new Runnable() { while(true) { @APE_If(“Battery.Level > 70%”) @APE_WaitUntil(“Network.Active”, MaxDelay=1200) @APE_Else() @APE_WaitUntil(“Network.Active”, MaxDelay=3600) attemptUpload(); } }; uploadThread.start();

APE: Annotated Programming for Energy-‐Efficiency

• A small and declara2ve annota2on language with a lightweight middleware run2me for Android •  Demarcate power-‐hungry code using annotaNons •  ExecuNon deferred unNl device enters state that minimizes cost of operaNon

•  Abstract away details of event/state monitoring

• Abstract model based on 2med automata •  Guides design of language and middleware

42

Outline •  IntroducNon • Approaches to Energy-‐Management • APE: Annotated Programming for Energy-‐efficiency •  Policy Model •  AnnotaNon Language •  RunNme Service

•  EvaluaNon •  Case Study: CiNSense •  Overhead and Power Savings

• Conclusion

43

The APE Policy Model • Restricted form of 5med automata:

•  Σ is a finite set of events •  𝑆 is a finite set of states •  𝑠0∈𝑆 is the start state •  𝑆𝐹⊆𝑆 is a set of accepNng states •  𝐶 is a finite set of clocks, and •  𝐸 is a transiNon funcNon

44

𝑇𝐴=(Σ, 𝑆, 𝑠0, 𝑆𝐹, 𝐶, 𝐸)

Alur, Rajeev, and David L. Dill. "A theory of Nmed automata." Theore5cal computer science 126.2 (1994): 183-‐235.

Typical Automata Can Be Expressed Textually

•  𝜎i is a Boolean expression consisNng of events and states •  𝑡i is a constraint on how long to wait for 𝜎i to be true •  𝑡𝑀𝑎𝑥𝐷𝑒𝑙𝑎𝑦 is an upper bound on the delay of O

•  Linear, concise nature is easy to express in wriNng •  Drove design of annotaNon language

45

𝑃=({(𝜎1,𝑡1), (𝜎2,𝑡2)…(𝜎𝑛,𝑡𝑛)},𝑡𝑀𝑎𝑥𝐷𝑒𝑙𝑎𝑦)

A Policy is a Sequence of Boolean Expressions and Timing Constraints • Defer the execuNon of an operaNon O unNl a finite sequence of states holds:

•  𝜎i is a Boolean expression consisNng of events and states

•  𝑡i is a constraint on how long to wait for 𝜎i to be true •  𝑡𝑀𝑎𝑥𝐷𝑒𝑙𝑎𝑦 is an upper bound on the delay of O

•  If 𝜎𝑖 is saNsfied before 𝑡𝑖, begin monitoring for 𝜎𝑖+1 •  Else, begin monitoring for 𝜎𝑖-‐1 again •  Execute O once 𝜎𝑛 is saNsfied or 𝑡𝑀𝑎𝑥𝐷𝑒𝑙𝑎𝑦 has passed

46

𝑃=({(𝜎1,𝑡1), (𝜎2,𝑡2)…(𝜎𝑛,𝑡𝑛)},𝑡𝑀𝑎𝑥𝐷𝑒𝑙𝑎𝑦)

Simplified Model: Limits Power for Conciseness

• Wait for driving to be detected a@er 30 seconds of con5nuous accelerometer movement

47

Not expressible in our model!

Allow for dynamic selecNon of policies at runNme

APE_If, APE_ElseIf, APE_Else: Condi*onals for Policy Selec*on

48

while(true) { @APE_If(“Battery.Level > 70%”) @APE_WaitUntil(“Net.Active”, MaxDelay=1800) @APE_Else() @APE_WaitUntil(“WiFi.Connected”, MaxDelay=3600) uploadSensorData(); }

@APE_WaitUntil(“MyPolicy, ({dataReady()},30)”, MaxDelay=3600) @APE_WaitUntil(“MyPolicy(inf,60), ({dataReady()},30)”, MaxDelay=3600)

APE_DefineTerm, APE_DefinePolicy: Enabling Reuse

49

@APE_DefineTerm(“MyTerm”, “Battery.Charging AND (WiFi.Connected OR Cell.4G)”)

@APE_WaitUntil(“{requestPending()} AND MyTerm”, MaxDelay=3600) @APE_DefinePolicy(“MyPolicy”, “(Display.Off,inf),(MyTerm,10)”)

while(true) { APEService.wait_until(1, 1800); uploadSensorData(); }

while(true) { @APE_WaitUntil(“WiFi.Connected”, MaxDelay=1800) uploadSensorData(); }

void onCreate(Bundle savedInstanceState) { super(savedInstanceState); … }

void onCreate(Bundle savedInstanceState) { super(savedInstanceState); APEService.register_expr(1, “WiFi.Connected”); … }

Annota*ons are Processed at Compile-‐*me into Run*me Requests

50

APE Run2me Service

The APE Service Performs Device State Monitoring on Behalf of Clients

51

Client App A

Android APIs

Client App B Client App C

Conclusion •  A small and declara2ve annota2on language with a lightweight middleware run2me for Android •  Demarcate power-‐hungry code using annotaNons •  ExecuNon deferred unNl device enters state that minimizes cost of operaNon

•  Abstract away details of event/state monitoring

• Manage trade-‐off between 2ming and energy

•  In a case study, a policy implemented with only six annota2ons dras2cally improved efficiency •  86% reducNon in lines of power-‐management code

52

Two Categories of Approaches to Energy-‐Management

•  Low-‐level Op2miza2ons: DVFS, Nckless kernels, radio scheduling, batched writes to flash, etc… •  Implemented in kernel, driver, or firmware •  Responsibility of the device and OS vendors

•  System-‐level Op2miza2ons: workload shaping, filtering, piggy-‐backing radio transmission, etc… •  Implemented at applicaNon level •  Responsibility of the applica2on developers

53

Documents

APE:AnAnnota*onLanguageand% …ochipara/presentations/ICSE2014.pdf · NetAct" OR 3G 4G. TheAPEPolicyModel ... 1 3 7 15 31 63 127 191 255 383 511 % StateExpressionLength Overhead"of"IPC"measured"to"be"1.71"ms"