A Mobility-aware Cross-edge Computation Offloading Framework for - - PowerPoint PPT Presentation

A Mobility-aware Cross-edge Computation Offloading Framework for Partitionable Applications

Hailiang Zhao12 Shuiguang Deng1 Cheng Zhang1 Wei Du2 Qiang He3 Jianwei Yin1

1Zhejiang University, Hangzhou, China 2Wuhan University of Technology, Wuhan, China 3Swinburne University of Technology, Melbourne, Australia

July 10, 2019

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 1 / 16

Outline

1

Introduction A Brief Introudction to Mobile Edge Computing (MEC) What is the Problem?

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 2 / 16

Outline

1

Introduction A Brief Introudction to Mobile Edge Computing (MEC) What is the Problem?

2

Cross-edge Computation Offloading System Model Problem Formulation Proposed Framework and Algorithms Experimental Evaluation

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 2 / 16

Introduction

Outline

1

Introduction A Brief Introudction to Mobile Edge Computing (MEC) What is the Problem?

2

Cross-edge Computation Offloading System Model Problem Formulation Proposed Framework and Algorithms Experimental Evaluation

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 3 / 16

Introduction A Brief Introudction to Mobile Edge Computing (MEC)

What is Mobile Edge Computing?

Mobile Edge Computing

Mobile Edge Computing (MEC) is a new computation paradigm:

1

depolyed at the network edge

2

use widespread wireless access network (Small-cell Base Station)

3

provide service and computing resource

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 4 / 16

Introduction A Brief Introudction to Mobile Edge Computing (MEC)

What’s it properties?

Edge site An edge site is a micro data center with applications depolyed, attached to a small-cell base station (SBS).

1 Heterogeneous edge sites 2 User mobility 3 Edge site selection and user profile handover 4 Overlapped signal coverage of SBSs (Corss-edge Collaboration!) 5 Partitionable applications (data stream) 6 Insufficient battery energy of mobile devices 7 ... ... Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 5 / 16

Introduction What is the Problem?

Motivation Scenario

270 90 180 315 45 225 135 User #1 User #2 User #3 time slot #1 time slot #2 time slot #3 time slot #4 time slot #5 time slot #6 time slot #1 time slot #2 time slot #3 time slot #4 time slot #5 time slot #6 time slot #1 time slot #2 time slot #3 time slot #4 time slot #5 time slot #6 partitioning & o partitioning & o

• Hailiang Zhao (Zhejiang University)

Cross-edge Computation Offloading July 10, 2019 6 / 16

Introduction What is the Problem?

Problem Definition For partitionable applications, how to make the offloading strategy with the minimum overall cost achieved?

Composition of overall cost

1 execution delay 2 task dropping penalty 3 collaboration cost

Energy Harvesting (EH) technology is adopted.

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 7 / 16

Cross-edge Computation Offloading

Outline

1

Introduction A Brief Introudction to Mobile Edge Computing (MEC) What is the Problem?

2

Cross-edge Computation Offloading System Model Problem Formulation Proposed Framework and Algorithms Experimental Evaluation

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 8 / 16

Cross-edge Computation Offloading System Model

System Model

1 Local execution latency evaluation 1

execution latency: τ lc

i ηl i/fi

2

energy consumption: ǫl

i κi · ηl if 2 i

2 Offloading latency evaluation 1

transmission delay: τ tx

i,j(t) µr

i

• j∈Mi(t) Ii,j(t) ·

1 Ri,j(t)

2

execution delay: τ rc

i,j(t) ηr

i

fj·

j∈Mi(t) Ii,j(t)

3

collaboration cost: ϕ ·

j∈Mi(t) Ii,j(t)

4

constraint: τd ≥ maxj∈Mi(t)

• τ tx

i,j(t) + τ rc i,j(t)

• + τ lc

i + ϕ · j∈Mi(t) Ii,j(t)

3 Battery energy level evaluation 1

envolution function: ψi(t + 1) = ψi(t) −

j∈Mi(t) ǫtx i,j(t) · Ii,j(t) − ǫl i + αi(t)

2

constraint: ǫl

i + j∈Mi(t) ǫtx i,j(t)Ii,j(t) ≤ ψi(t)

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 9 / 16

Cross-edge Computation Offloading Problem Formulation

Problem Formulation

A non-convex optimization problem P1 : min

∀i,Ii(t),αi(t)

lim

T→∞

1 T

T−1

• t=0

E

i∈N

C(Ii(t))

• ,

with several constraints. C(Ii(t))

• max

j∈Mi(t):Ii,j′(t)=1

• τ tx

i,j(t) + τ rc i,j(t)

• +

τ lc

i + ϕ ·

• j∈Mi(t)

Ii,j(t) + ̺i · Di(t)

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 10 / 16

Cross-edge Computation Offloading Proposed Framework and Algorithms

Proposed Framework

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 11 / 16

Cross-edge Computation Offloading Proposed Framework and Algorithms

Proposed Algorithms

The CCO algorithm

1 Lyapunov optimization (drift-plus-penalty)

P2 : min

∀i,Ii(t),αi(t) ∆up V (Θ(t)),

with several constraints. ∆up

V (Θ(t))

• N
• i=1

ψ′

i(t)

• αi(t) − ǫl

i − M

• j=1

ǫtx

i,j(t)Ii,j(t)

• +

V

N

• i=1

C(Ii(t)) + C

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 12 / 16

Cross-edge Computation Offloading Proposed Framework and Algorithms

Proposed Algorithms

The CCO algorithm Algorithm 1 Cross-edge Computation Offloading (CCO)

1: At the beginning of the tth time slot, obtain i.i.d.

random events A(t), Eh(t) [Eh

1 (t), ..., Eh N(t)] and channel state information.

2: ∀i ∈ N, decide I⋆

i (t), α⋆ i (t) by solving the deterministic problem P2.

3: ∀i ∈ N, update the battery energy level ψi(t). 4: t ← t + 1.

1 optimal energy harvesting: α⋆

i (t)

2 optimal edge site selection: I⋆

i (t)

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 13 / 16

Cross-edge Computation Offloading Experimental Evaluation

Benchmark Policies

1

Random Selection (RS)

2

Greedy Selection on Communication (GSC1)

3

Greedy Selection on Computation (GSC2) Parameter settings

Parameter Value Parameter Value τd 2 ms ̺i 2 ms ϕ 0.02 ms ρi 0.6 µl

i

100 bits µr

i

3000 bits fi 1.5 GHz fj 32 GHz κi 10−28 ψsafe

i

40 mJ N max

j

5 ω 1.5/

i∈Nj(t) Ii,j(t) GHz

̟0 10−13 W ptx

i

1 W g0 10−4 Emax

i,h

4.8 × 10−4 J

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 14 / 16

Cross-edge Computation Offloading Experimental Evaluation

Optimality and stability

25 50 75 100 125 150 175 200 time slot 0.11 0.12 0.13 0.14 0.15 0.16 Average cost of all mobile devices

Overall costs in different time slots

CCO algorithm RS GSC1 GSC2

Figure: Average cost of mobile devices.

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 15 / 16

Cross-edge Computation Offloading Experimental Evaluation

Optimality and stability

25 50 75 100 125 150 175 200 time slot 0.0150 0.0175 0.0200 0.0225 0.0250 0.0275 0.0300 0.0325 Battery energy levels

Average battery levels in different time slots

CCO algorithm RS GSC1 GSC2

Figure: Average battery energy level of mobile devices.

Hailiang Zhao (Zhejiang University) Cross-edge Computation Offloading July 10, 2019 16 / 16