Ready, Set, Go: Coalesced Offloading from Mobile Devices to the - - PowerPoint PPT Presentation

Ready, Set, Go: Coalesced Offloading from Mobile Devices to the Cloud

Liyao Xiang, Shiwen Ye, Yuan Feng, Baochun Li, Bo Li Department of Electrical and Computer Engineering University of Toronto May 1st, 2014

Remote execution

2

External Storage Application Server Application Server

3

Smartphone

Profiler Solver Appli- cation

Application Server

Cloud

Appli- cation Solver Profiler

Code offloading

Tail time phenomenon

• When multiple applications send their
• ffloading requests without

coordination, network interface enters at high-power state at arbitrary times.

4

Coalesced Offloading

5

Smartphone

Application Server

Profiler Solver

Offloading Requests Coordination Service

Appli- cation App Solver Profiler App Solver Profiler

Application Server

Cloud

Offloading Requests Coordination Service

Coalesced Offloading

6

Time

t1

Time Power State Power State (a) Before bundling: (b) After bundling:

requests

• f app 1

requests

• f app 2

t2 t3 t4 t5 t6 t7 t2(t1') t3 t5(t4') t7(t6')

Problem Formulation

• Assume that M applications, generating requests at

. The requests are granted at .

7

a1, a2, ... g1, g2, ...

Tail Time T Power State Time

a3(g1) = t1

High-power State Low-power State

s1 a2 a1 a5(g3) a4(g2) a6 a7 a8(g4) = t2 a9(g5)

latency(1)

Problem Formulation

• Energy cost function
• Latency cost function
• The joint optimization problem is as follows:

8

min fcost = X

j

min{gj − gj−1, T} + α X

j

X

ai s.t. gj−1≤ai≤gj

(gj − ai) X

j

min{gj − gj−1, T} ' X

j

X

ai s.t. gj−1≤ai≤gj

(gj − ai) =

How to solve the problem?

9

RSG Solutions

10

• Optimal offline algorithm:
• With the arrival time sequence

known a priori.

• Online algorithms.
• Without a priori knowledge of the arrival time

sequence.

a1, a2, . . . , an

RSG Offline Solution

• For request ,
• For Combinations of binary transmission sequence,

we should:

• The problem is transformed from continuous-time to

discrete-time formulation.

11

ai f i

cost =

( min{ai − gprev, T}, if granted, α(gnext − ai), if delayed. 2n min fcost =

n

X

i=1

f i

cost

What if we don’t know the entire input sequence?

12

Our Results

• Algorithm is 2-competitive.
• The competitive ratio between the

expected cost incurred by and the

• ptimal cost is .
• RSG Online Algorithm have the optimal

competitive ratio.

13

A1 A e/(e − 1)

Performance Evaluation

• Measuring the Tail Time (on iPhone 3GS, Bell Mobility 3G network)
• Transmitting successive packets of equal size with constant

transmission intervals.

• Model-driven Simulations
• Simulating the timing of multiple offloading requests from

several simultaneously running applications.

• Real-world Experiments

14

15

Experiment Results

Random Bursty Stable 0.02 0.03 0.04 0.05 0.06 0.07 0.08 Unit Time Energy Cost (mAh) Naive Deterministic Randomized Offline

Energy consumption with different types of requests

0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 α

Unit Time Energy Cost (mAh)

Naive Offline Randomized Deterministic

Energy consumption with varying alpha

Experiment Results

16 Naive Online

Time (s) Request Grants Rubik Solver Email Chat

Real requests on mobile device w/o RSG solutions

50 100 150 200 250 4040 4060 4080 4100 4120 4140

Raw Battery Voltage (mV)

50 100 150 200 250 4040 4060 4080 4100 4120 4140

Time (s) Raw Battery Voltage (mV)

Battery Voltage Change on mobile device w/o RSG solutions

Conclusions

• By bundling the offloading requests of

multiple applications, we achieve greater energy savings while maintaining satisfactory performance.

• The RSG online algorithm achieves the

best possible competitive ratio.

17

Thank you.

18