Network SoftwarizationStays to Become A Reality...Network SoftwarizationStays to Become A Reality Aki Nakao Professor, Department Chair of GSII The University of Tokyo Executive Senior

Network Softwarization Stays to Become A Reality

Aki NakaoProfessor, Department Chair of GSII

The University of Tokyo

Executive Senior Fellow, Keidanren

Chairman, 5GMF Network Architecture Committee2018/12/11

All Rights Reserved by Akihiro Nakao, 2018

ETSI/OSA Workshop

1


UTokyo has been a member of ETSI and 3GPP




Open Implementationand

Standardization

5

Open Source Implementation is a great step toward standardization


What is Softwarized RAN for?

• Edge Computing Consolidation•Data Analytics• Slicing for eMBB/URLLC Applications

6


Our 5G Field Trials for eMBB/URLLC

7



From Autonomous Driving to Coordinated Driving

https://newsroom.intel.com/newsroom/wp-content/uploads/sites/11/2017/05/passenger-economy.pdf

INTEL says the Passenger Economy represents a 7-trillion-USD global opportunity in 2050.

According to the report of NHTSA in 2015, about 94% of traffic accidents had been caused by drivers (human errors)

9


Coordinate Driving:

Can we get rid of traffic signals?

10

All Rights Reserved by Akihiro Nakao, 2018https://www.bsfilms.me「Rush Hour」

11


Autonomous Intersection Management (AIM)

Intersection Manager (IM)

Request Confirm

Request

Reject

http://www.cs.utexas.edu/~aim/

Location / Speed Of Vehicles（URLLC）

Low Latency Feedback Is necessary

Ultra Reliable Low Latency and Edge Computing are Mandatory12

Coordinated Driving via MEC• URLLC (Ultra Reliable and Low Latency Communication) Applications• MEC: Multi-Access/Mobile Edge Computing• Harnessing edge servers• Ultra low end-to-end delay• Insufficient computational resources

Internet

GWOperator NWEdge server

Cloud

eNBKengo Sasaki and Naoya Suzuki and Satoshi Makido and Akihiro Nakao Vehicle Control System Coordinated between Cloud and Mobile Edge Computing, IEEE SICE (2016 Kengo Sasaki and Naoya Suzuki and Satoshi Makido and Akihiro Nakao, “Layered Vehicle Control Sys- tem Coordinated between Multiple Edge Servers,” 3rd IEEE Conference on Network Softwarization (IEEE NetSoft) 2017

13

NTT Docomo x UTokyo Collaborationon Cooperative Driving

oPress Released on Nov. 28th , 2018oNTT Docomo Open House Dec. 6th and 7th, 2018

Experimetal Settings 5G Equipment (4.5GHz)

Cars at Intersection Collision Avoidance (Before) Collision Avoidance(After)14

• Press Released on Nov. 28th , 2018

• NTT Docomo Open House Dec. 6th and 7th, 2018

NTT Docomo x Utokyo Collaborationon Cooperative Driving

Driving Trajectory

Ultrasonic Sensors

5GURLLC

15

NTT Docomo Open House Dec, 6th and 7th

16


0102030405060708090

100

0ms 50ms 100ms 150ms 200ms 250ms 300ms

Micro-benchmarkDelay vs. Course Out Ratio

If e2e delay is larger than 150ms, more than 40% of the entire trajectory is course out

Edge server control Delay

Cou

rse

out r

atio

[%]

From APNOMS2016

17



Japan's First Real Time 4K Video Transmission using 5G DroneJune 2018

4K/8K Realtime Drone Surveillance using 5G Cellular

19




4K RealTime Video Transmission

22


Stitching 4K Video Feeds From 6 Fisheye-lens Cameras

23


Cycling Shimanami 2018

24


Cycling Shimanami 2018

25








1G

>10G

100M

100

1,000

10,000

101

50500

x1000 <

Typical User Throughput (bps)

Latency (ms)(RAN R.T. delay)

Mobility (km/h)

Number of Connected Users

Energy Saving

1

High Energy Saving

1

Capacity (bps/km2)

1/3<

High Mobility

High Capacity

High Number of Connected Users

5G Mobile Key Performance Indicators (KPI)

Ultra Reliable andLow Latency Communication(URLLC)

Real-Time Object Recognition via Deep Neural Network

Small form factor GPU for DNN in Drones

Jetson TX2

Real-Time Object Recognition via Deep Neural Network


Real-Time Object Recognition via Deep Neural Network from Drones

37


Real-Time Object Recognition via Deep Neural Network from Drones

38


https://www.youtube.com/watch?v=cxyPUmeXwjw&list=PLUcnRyRXHsDhqzXSi-3yoo1KifSVcv-bqYoloV2 with darknet pretrained weights

The Same Recognition Model Applied to a YouTube Drone Video

39


The Recognition Model Trained With Stanford Pedestrian Data

http://cvgl.stanford.edu/projects/uav_data/40


RAN Slicing Comes Into Play!

41

1G

>10G

100M

100

1,000

10,000

101

50500

x1000 <

Typical User Throughput (bps)

Latency (ms)(RAN R.T. delay)

Mobility (km/h)

Number of Connected Users

Energy Saving

1

High Energy Saving

1

Capacity (bps/km2)

1/3<

High Mobility

High Capacity

High Number of Connected Users

5G Mobile Key Performance Indicators (KPI)

Enhanced Mobile Broadband(eMBB)

Ultra Reliable andLow Latency Communication(URLLC)

4K 360 degree Video Streaming

5G

URLLC Slice3D Mapping InformationObject Recognition by DNN

Cloud

Implication : Network Slicing for URLLC and eMBB

Edge

Network Slicing

eMBB Slice

Application based Traffic Classification(RAN Slicing, Machine Learning)

43


RAN Slicing using OAI

44


Application SpecificRAN Slicing extending FlexRAN

• Each RAN slice has its own RBs and MAC schedulers• Assign UE to RAN slice at a granularity of UE 45


Experimental Results

DL: 2.11MbpsUL: 0.94Mbps




Two App slices: SpeedTest and DefaultSpeedTest Slice is with different percentage of RBs



48

3 RBs

6 RBs

9 RBs

12 RBs

15 RBs

18 RBs


0

2

4

6

8

10

12

14

3 6 9 12 15 18

Throughput (Mbps)

Resource Blocks (RBs) assigned to SpeedTest Slice

Throughput of SpeedTest app

DL UL

The throughput of both UL and DL is proportional to the number of assigned RBs

Throughput Evaluation

49


Experimental Results

50RBs (100%)

36RBs(75%)

24RBs (50%)

12RBs (25%)





Two App slices: SpeedTest and DefaultSpeedTest Slice is with different percentage of RBs

50


Throughput Evaluation

0

10

20

30

40

25% 50% 75% 100%

Throughput of SpeedTest app

DL UL

# Proportion of RBs of SpeedTest slice

Throughput (Mbps)

51


Experiments (2)• App Cloner: SpeedTest, SpeedTesu• SpeedTesu is a clone of SpeedTest, where everything is the same expect appname

SpeedTest SpeedTesu

52


Experiments (2)SpeedTest(25% RBs)

SpeedTesu(50% RBs)

• With App-specific UE Slicing, different Applications can get different RAN resources even they are on the same UE.

54


Conclusion

Softwarization stays to become a reality Complex data processing and flexible

network operation will be enabled in programmable softwarized infrastructure

55

Two Implications:

• Commercial Use Cases Initiated in Rural Arease.g. Local Governments, Agriculture/Fishery IoT

• Academia is now empowered with CRAZY idease.g. RAN Slicing, Context Slicing, etc.

Documents

Network SoftwarizationStays to Become A Reality...Network SoftwarizationStays to Become A Reality Aki Nakao Professor, Department Chair of GSII The University of Tokyo Executive Senior