Learning-Assisted Network Slicing for Diverse Applications in 5G

Tao Han

The Department of Electrical and Computer Engineering

The University of North Carolina at Charlotte, NC, United States

Tao.Han@uncc.edu

https://webpages.uncc.edu/than3/index.html

What are “killer” applications for 5G?

2

Tourist & Navigation

Smart Industry

Education

GamingAutonomous Vehicle

Battlefield

➢Diverse resource requirements from multiple domains: ▪ Radio Access Networks ▪ Transportation▪ Computing

Network Slicing: End-to-End Customization

3

* Guan, W., Wen, X., Wang, L., Lu, Z. and Shen, Y., 2018. A service-oriented deployment policy of end-to-end network slicing based on complex network theory. IEEE Access, 6, pp.19691-19701.

Isolation v.s. Multiplexing

4

Segment 1 Segment 2 Segment 3Resources:

Slices:

Tim

e

01

2… Can we provide isolation and

allow multiplexing?

What is the “Capacity” formula?

➢ The Shannon–Hartley theorem

𝐶 = 𝐵 𝑙𝑜𝑔2(1 + 𝑆𝑁𝑅)

5

RAN Transport Edge/Cloud Computing

20 msGood

Better

Best

Good

Better

Best

Good

Better

Best

200 ms Good

Better

Best

Good

Better

Best

Good

Better

Best

Learning-Assisted Dynamic Network Slicing

1. Slice tenant requests a network slice

2. Admission control of slice requests

3. Orchestrator allocates the multiple domain resources (virtual) for all admitted network slices

4. Each network slice allocates resources (virtual) to its users

5. The virtual resource allocations of all the users are informed to hypervisor

6. The hypervisor maps the virtual resources to physical resources to maximize the efficiency of physical resources

6

Qiang Liu and Tao Han, "VirtualEdge: Multi-Domain Resource Orchestration and Virtualization in Cellular Edge Computing", IEEE International Conference on Distributed Computing Systems (ICDCS) 2019.

Admission

Control

Slice Request

Slice Request

Multi-Domain Resource Orchestrator1

Multi-Domain Resource Hypervisor 2

Virtual Resource Virtual Resource

User to Virtual

Resource Mapping

User to Virtual

Resource Mapping

Virtual to Physical Resource Mapping

Slice N

Utility Update

Slice N

Slice 1

Utility Update

Slice 1

Multi-Domain

Resource

Scheduling

Multi-Domain

Resource

Scheduling

System Implementation

GPU

vNode #1

vNode #2

Multi-Domain Resource Orchestrator

UE#1@vNode#1

UE#2@vNode#1

UE#1@vNode#2

UE#2@vNode#2eNodeB GPU

Cellular Edge Computing Node

eNB

vNode #1

vNode #2

7

The system is developed and implemented based on the OpenAirInterface (OAI) LTE and CUDA GPU computing platforms

Radio Resource Hypervisor

➢ Managing the MAC layer user scheduling and resource allocation (physical resource blocks (PRBs) in LTE network).

8

PRB#1 PRB#2 PRB#3 PRB#4 PRB#5 PRB#6

Sub-Carriers

vNode#1,

RR=360kHz

vNode#2

RR=540kHz

vNode#3

RR=180kHz

1 S

ub

frame

#1

Users B

uffers

RLC and upper layers

MAC layer

Physical layers

The Virtual-to-Physical Mapping

#2 #1 #2 #1

Computing Resource Hypervisor

➢ Managing the dispatch of kernel functions (Token-based)

9

name threads parameters

Kernel function inCUDA programing:

Kernel 1 (10k threads)

Kernel 2(1k threads)

Kernel 3(5k threads)

Kernel NCal

ling kernels

asyn

chro

no

usl

y

Kernel 1(10k threads)

Kernel 2(1k threads)

Kernel 3(5k threads)

Executing kernels serially in GPU

CPU side

GPU side

Resource Hypervisor: Computing

➢ Methodology: Managing the dispatch of kernel functions (Token-based)

10

name threads parameters

Kernel function inCUDA programing:

Kernel 1 (10k threads)

Kernel 2(1k threads)

Kernel 3(5k threads)

Kernel NCal

ling kernels

asyn

chro

no

usl

y

Kernel 1(10k threads)

Kernel 2(1k threads)

Kernel 3(5k threads)

Executing kernels serially in GPU

CPU side

GPU side

Kernel 1(10k threads)

Kernel 2(1k threads)

Kernel 3(5k threads)

Command Queue

Token

Distributed Resource Orchestration

➢ Considering multiple eNodeBs and computing servers

11

Radio Access Network Edge Servers

Qiang Liu and Tao Han, “DIRECT: Distributed Cross-Domain Resource Orchestration in Cellular Edge Computing”, ACM International Symposium on Mobile Ad Hoc Networking and Computing(MOBIHOC) 2019.

Algorithm Overview

➢ Controller side: ▪Updating the dual variables and optimize the auxiliary

variable Z (convex problem)

➢ Node side: ▪Optimize the resource allocation X

12

System Overview

➢ DIRECT controller: Coordinate the resource allocations to slices across edge nodes (control-side algorithm)

➢ DIRECT agents in edge nodes: Allocate resources to slices using a learning-based algorithm (edge-side algorithm)

13

➢ Slice Orchestrator: Dynamically orchestrate virtual network resources to slices across the network

System Overview

14

➢ Resource Hypervisor: map virtual resource to physical resources

Radio Resource Hypervisor

Computing Resource Hypervisor

System Implementation

➢ Hardware Details: ▪OpenAirInterface (OAI): 2x USRP B210 SDR boards, 2x

eNodeB computers, 1x Core network▪CUDA GPU computing platform: 2x NVIDIA GTX 1080Ti,

CUDA 8.0▪ Mobile users: 4x Huawei dongle E2273, 4x Linux computers

15

GPU 1

GPU 2

Core

Network

eNodeB 1

UE 1@Slice 1

UE 2@Slice 3

UE 3@Slice 2

UE 4@Slice 3

Edge Node 1 with DIRECT Agent

Edge Node 2 with DIRECT Agent

Th

e D

IRE

CT

Contr

oll

er

eNodeB 2

Experiment Results

➢ DIRECT is aware of the traffic load

➢ DIRECT learns the needs of multi-domain resources of slices

16

Res

ou

rce

uti

liza

tion

Uplink Computing

Slice 1

10%

20%

30%

40%

50%

60%

Downlink

Slice 2 Slice 30%

Res

ou

rce

uti

liza

tion

Slice 10%

25%

50%

75%

100%

Slice 2 Slice 3

Edge node 1 Edge node 2

(b)(a)

application MAR MAR VAS

Learn the slice traffic on edges

Learn the resource demand of application

Experiment Results

➢ DIRECT converges in a few iterations.➢ DIRECT reduces about 21% system latency as compared

to Static.➢ DIRECT agents can learn the optimal resource allocations

to network slices.

17

0 2 4 6 8 102.0

2.5

3.0

3.5

4.0

4.5

5.0

Iteration

Syst

em-l

aten

cy (

s)

Static

DIRECT

2 4 6 8 10Iteration

Gap

0.0

0.1

0.2

0.3

0.4

0

(a) (b)Iteration

Edg

e-la

tency

(s)

0 5 10 15 20 25 30

1.0

1.5

2.0

2.5

Edge node 2

Edge node 1

(c)

21%

61%

18

Tao Han

The University of North Carolina at Charlotte

Tel #:704-687-8406, Email: Tao.Han@uncc.edu

AI