An Online Learning-Based Task Offloading Framework for 5G Small Cell - - PowerPoint PPT Presentation

An Online Learning-Based Task Offloading Framework for 5G Small Cell Networks

ICPP2020 1

2 ICPP2020

Outline

 Background & Motivation  System Model  Algorithm Design  Analysis & Performance  Conclusion

3 ICPP2020

Background & Motivation - 5G

5G technology: the fifth-generation mobile communication technology

• Higher transmission rate
• Faster speed
• Larger number of connections

4 ICPP2020

Background & Motivation - small cell node

Small cell node (SCN): fundamental element of 5G network

• Low-power, short-range
• Cover small geographical areas
• Millimeter-Wave
• Closer to wireless devices

Process larger amount of data at faster speeds

5 ICPP2020

Background & Motivation - small cell network

• SCN
• Improve coverage and capacity
• Better and faster connectivity
• Macrocell
• Wide coverage
• Further away from the users
• Communicate with SCNs
• Fig. An illustration of small cell network

6 ICPP2020

Background & Motivation

• 5G boosts demand for new services
• Security surveillance
• Virtual reality
• Automatic driving
• …….
• Characteristic of services
• Huge amount of data
• Delay sensitive
• High computing requirements

Small cells equipped with edge servers represent a competitive solution for mobile task offloading

7 ICPP2020

Background & Motivation

• Features of offloading tasks from wireless devices (WDs) in small cell network
• Limited connections
• Finite computation resources
• Significant uncertainties in the task offloading process, such as blockage
• A fundamental problem

Given limited computation and communication resources, how to select computing tasks to maximize effective reward?

8 ICPP2020

Outline

 Background & Motivation  System Model  Algorithm Design  Analysis & Performance  Conclusion

9 ICPP2020

System Model - overview

MBS:

• Controls and prioritizes offloading

task to SCNs SCN：

• 𝑁 SCNs
• Limited coverage
• Equipped with computing server

Task：

• Provide task context information
• May be covered by multiple small

cells

10 ICPP2020

System Model - goal

• Random process in task offloading:
• Reward for SCN 𝑛 to complete task 𝑗 at time 𝑢

𝑉 𝑛, 𝜚𝑗, 𝑢

• Likelihood for SCN 𝑛 to complete 𝑗 at time 𝑢

𝑊(𝑛, 𝜚𝑗, 𝑢)

• Resource consumption 𝑅(𝑛, 𝜚𝑗, 𝑢)
• Our goal:
• Compound reward: the effective reward per unit resource for SCN 𝑛 to

complete the task with context 𝜚 at time 𝑢 𝐻 𝑛, 𝜚𝑗, 𝑢 = 𝑉 ∗ 𝑊/𝑅

• Maximize the total compound reward under system constraints

11 ICPP2020

System Model - constraints

• The number of tasks accepted by each SCN does not exceed its

communication capacity

• Each task is not repetitively offloaded by multiple SCNs
• Number of successfully processed tasks by each SCN at a time slot
• Resource capacity

12 ICPP2020

Outline

 Background & Motivation  System Model  Algorithm Design  Analysis & Performance  Conclusion

13 ICPP2020

Algorithm Design - difficulties

1. Uncertain and stochastic environment: reward, likelihood, resource consumption 2. The balance between maximizing the total compound reward and satisfying the system constraints 3. Enumerating all possible sets and selecting the optimal one leads to large search space. How to avoid combinatorial explosion? 4. How to guarantee tasks are not repeatedly offloaded?

14 ICPP2020

Algorithm Design - technique 1

• Multi-armed bandit (MAB)
• A model for sequential decision problems with

exploration-exploitation tradeoff

• Some bandit machines (arms), and drawing an

arm yields a reward.

• Lagrangian multipliers
• General MAB framework does not consider any constraints
• Introduce adjustable Lagrangian multipliers, to balance reward and constraint

violation

• Construct a new regret function

𝑍 = 𝑆 𝑛, 𝑢 + 𝜇1 𝑛, 𝑢 ∗ 𝑊

1 𝑛, 𝑢 2 + 𝜇2 𝑛, 𝑢 ∗ 𝑊 2 𝑛, 𝑢 2

15 ICPP2020

Algorithm Design - technique 2

Oracle total compound reward - our algorithm total compound reward

16 ICPP2020

Algorithm Design - technique 3

• Tailored Contextual MAB
• Each task comes with its context
• Massive contexts to be learned
• Divide task context space into small hypercubes of similar contexts
• Maintain a weight for each hypercube, based on the regret 𝑍
• For each SCN, estimate the selection probability of each task

17 ICPP2020

Algorithm Design - technique 4

• Weighted bipartite graph
• A task may be covered by multiple SCNs
• Coordinate multiple SCNs for task offloading
• Construct weighted bipartite graph, based on all

tasks’selection probabilities

• Greedy manner

S1 S2 k1 k2 k3 k4 0.7 0.9 0.6 0.4 0.7 0.8

18 ICPP2020

Algorithm Design - framework

19 ICPP2020

Outline

 Background & Motivation  System Model  Algorithm Design  Analysis & Performance  Conclusion

20 ICPP2020

Analysis & Performance - theoretical analysis

• Regret = total reward of oracle – total reward of our algorithm
• Violations of system constraints
• Regret and violations are all sub-linear with respect to timespan 𝑈
• Converges to the optimal task offloading decisions over time
• Asymptotically optimal performance when 𝑈 is sufficiently large

21 ICPP2020

Analysis & Performance - performance

• Simulation setup
• 10000 time slots
• 30 SCNs connected to a MBS
• 35-100 tasks in the coverage area of each SCN
• each SCN can simultaneously support up to 20 wireless device
• Reward and likelihood are normalized and uniformly distributed in [0, 1]

22 ICPP2020

Analysis & Performance - performance

• Performance Metrics
• Cumulative compound reward (violation)
• Per-time-slot compound reward (violation)
• Performance ratio

= cumulative compound reward / (cumulative violation1 + cumulative violation2)

23 ICPP2020

Analysis & Performance - performance

Fig1 Fig2 Fig3 Fig4

24 ICPP2020

Outline

 Background & Motivation  System Model  Algorithm Design  Analysis & Performance  Conclusion

25 ICPP2020

Conclusion

• Study task offloading in 5G small cell networks
• Propose an online learning-based solution framework
• Leverage MAB technique to learn the best selection strategy, while considering

resource capacity constraints and QoS requirement

• The efficiency is verified by both theoretical analysis and simulation studies
• Achieve sub-linear bounds for both regret and violations

Thanks for listening！

26 ICPP2020