5g-miedge · 2019. 9. 26. · 2 5g-miedge page 2 the document is proprietary of the 5g-miedge...
TRANSCRIPT
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date : July 2019 Public Deliverable
5G-MiEdge Page 1
5G-MiEdge Millimeter-wave Edge Cloud as an Enabler for 5G Ecosystem
EU Contract No. EUJ-01-2016-723171
Contractual date: M36
Actual date: M36
Authors: See list
Work package: D4.3 MiEdge field trials integrated in 5G-Berlin Testbed toward Tokyo
Olympic 2020
Security: Public
Nature: Report
Version: 1.0
Number of pages: 51
Abstract
This deliverable reports on feasibility and availability of all key 5G MiEdge network components and
algorithms and their measured performance after implementation.
Keywords
millimeter-wave, millimeter-wave access, millimeter-wave backhaul, edge cloud, edge computing, liquid
ran c-plane, user/application centric orchestration, sdn, software-defined network, multi-rat
All rights reserved.
2
5G-MiEdge Page 2
The document is proprietary of the 5G-MiEdge consortium members. No copy or distribution, in
any form or by any means, is allowed without the prior written agreement of the owner of the
property rights.
This document reflects only the authors’ view. The European Community is not liable for any use
hat may be made of the information contained herein.
Authors
Fraunhofer Heinrich Hertz
Institute
Konstantin Koslowski
Thomas Haustein
Julian Daube
Tokyo Institute of Technology Kei Sakaguchi
Gia Khanh Tran
Hiroaki Nishiuchi
Makoto Nakamura
Yu Tao
Intel Valerio Frascolla [email protected]
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date : July 2019 Public Deliverable
5G-MiEdge Page 3
Table of contents
Abbreviations .................................................................................................................. 5
Executive Summary ........................................................................................................ 7
1 Introduction ............................................................................................................. 8
2 Testbed Architecture and Selected Scenarios ....................................................... 9
2.1 Testbed Architecture ......................................................................................... 9
2.2 Selected scenarios for evaluation ..................................................................... 9
2.2.1 Dynamic Backhaul Reconfiguration .................................................... 9
2.2.2 Dynamic Crowd.................................................................................. 12
2.2.3 Automated Driving ............................................................................. 12
2.3 Targeted Features and Details of Scenario-specific implementation ............. 13
2.3.1 Dynamic Backhaul Reconfiguration .................................................. 13
2.3.2 Dynamic Crowd.................................................................................. 14
2.3.3 Autonomous Driving .......................................................................... 15
3 Deployment of Testbed and Operational Testing ............................................... 17
3.1 Dynamic Backhaul Reconfiguration .............................................................. 17
3.2 Dynamic Crowd ............................................................................................. 18
3.2.1 Outdoor testbed architecture............................................................... 20
3.2.2 Definition of KPIs .............................................................................. 22
3.2.3 Results ................................................................................................ 23
3.3 Autonomous Driving ...................................................................................... 24
4 Measurements and Evaluation Results ............................................................... 27
4.1 Dynamic Backhaul Reconfiguration .............................................................. 27
4.1.1 Indoor testbed architecture ................................................................. 27
4.1.2 Definition of KPIs .............................................................................. 28
4.2 Dynamic Crowd ............................................................................................. 37
4.2.1 Results on Prefetching ........................................................................ 37
4.2.2 Results on Route Multiplexing ........................................................... 39
4.2.3 Results on Integration with Internet and Wi-Fi .................................. 40
4.2.4 Results on Multiple UEs ..................................................................... 41
4.2.5 Results on Blocking mmWave access ................................................ 42
4
5G-MiEdge Page 4
4.3 Autonomous Driving ...................................................................................... 44
4.4 Integration of 5G!Pagoda and 5G-MiEdge testbeds ...................................... 45
5 Final Demo at EuCNC .......................................................................................... 47
6 Summary ................................................................................................................ 50
7 References .............................................................................................................. 51
5
5G-MiEdge Page 5
Abbreviations
Acronym Description
AP Access Point
BS Base Station
C plane Control Plane
EC Edge Computing
EU European Union
HD High Definition
IT Information Technologies
JP Japan
KPI Key Performance Indicator
LiDAR Light Detection and Ranging
LOS Line of sight
MBS Macro base station
MEC Multi-access Edge Computing
mmWave Millimeter-wave
OBU On-Board Unit
ODC Outdoor Dynamic Crowd
OVS Open vSwitch
RAN Radio Access Network
RAT Radio Access Technology
RF Radio frequency
RRM Radio Resource Management
RSSI Received Signal Strength Indicator
RSU Road-Side Unit
SC-BS Small Cell Base Station
SC Small Cell
SDN Software-Defined Network
SOCRA Software-Defined Small Cell Radio Access Network
UE User Equipment
U-plane User Plane
V2V Vehicle to Vehicle
6
5G-MiEdge Page 6
V2X Vehicle to Everything
WP Work Package
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 7
Executive Summary
The EU-Japan funded research project 5G-MiEdge (Millimeter-wave Edge cloud as
an enabler for 5G ecosystem) envisions to design, develop and demonstrate an
innovative 5G architecture, which manages to smartly and smoothly combine
millimeter-wave (mmWave) access/backhauling with edge computing (EC). This
document is the final deliverable in WP4, which runs from month 7 to 36 of the 36-
month lifetime of the project.
WP4 is dedicated to the evaluation of the 5G system performance enhanced by the
MiEdge concepts using system level simulation tools and real world field-tests in the
5G Berlin and Tokyo testbeds for indoor and outdoor. The work includes numerical
system simulations on high performance computing clusters to capture system relevant
KPIs and effects in repeated and controlled modelling environments representing
typical use cases under investigation. After all necessary steps of development,
integration and testing of the key MiEdge network components are completed; we
conducted field trials for specific use cases and scenarios, which were previously
evaluated by system level simulations. The Over-the-Air tests were conducted in the
5G-MiEdge testbeds in Berlin and Tokyo under real world environment as feasibility
proof of concept and evaluation of the key achievements of the 5G MiEdge project.
Joint demonstrations between EU and Japan are the central scope of WP4.
This deliverable begins with an introduction of the used architecture, the selected
scenarios and the features, which are to be evaluated. Then the deployment and
operation testing of the testbeds is explained, before presenting the measurement
results.
Another section describes our final presentation on the EuCNC 2019 in Valencia,
Spain and virtual joint measurements with the 5G-Pagoda! Project.
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 8
1 Introduction
This deliverable is the final release of WP4 and describes the outcome of the work
done in Task 4.2 ‘Development of common/joint 5G MiEdge Testbed’ and Task 4.3 ‘:
Field trials toward Tokyo Olympic 2020’. While Task 4.2 was already detailed in the
previous deliverable D4.2 [D4.2], there are some notable changes related to the main
subject of this deliverable, mainly worked on by Task 4.3.
When the project was planned in 2016, we wanted to put our focus on the Tokyo
Olympics 2020, but this focus shifted during the project runtime. We still cover related
scenarios like the stadium and demonstrate that the key enabling technologies
developed in this project are ready for large events like the Olympics, but also general
5G deployments like the outdoor dynamic crowd. Lately the autonomous driving
scenario gained a lot of interest, from researchers as well as the automotive industry
and especially when combining mmWave and EC elements to achieve low latency and
high data rate Vehicle-to-Vehicle (V2V) and Vehicle-to-Everything (V2X)
communication.
In Section 2 “Testbed Architecture and Selected Scenarios” starts describing the
architecture proposed by the project, the selected scenarios and the features targeted
for evaluation, in order to provide a detailed evaluation of the outdoor capabilities of
the 5G-MiEdge testbeds in Berlin and Tokyo.
The following “Section 3 Deployment of Testbed and Operational Testing”, the exact
setup for evaluation of each of the selected scenarios is detailed.
Section 4 “Measurements and Evaluation Results” presents the measured results and
explains the difference between our baseline setup and the evaluated scenarios.
Moreover, an analysis is provided about when it is advantageous using a dedicated
network slice over the public network for intercontinental communication.
Section 5 “Final Demo at EuCNC” displays details of our final presentation at the
EuCNC 2019 in Valencia, Spain.
Finally, Section 6 “Summary“ concludes this deliverable.
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 9
2 Testbed Architecture and Selected Scenarios
In deliverable D4.2 [D4.2], we introduced the 5G-MiEdge testbeds, which consist
mostly of off-the-shelf hardware that we enhanced with specifically developed custom
components and runs open source software, which we adapted to fit the project
requirements. Based on the architecture proposed by the 5G-MiEdge project, we
created a viable testing environment for the novel developed ideas and concepts.
The testbed architecture can be adapted to support our introduced use-cases and
scenarios, verify the simulation results and highlight the improvements over current
state of the art deployments.
2.1 Testbed Architecture
Figure 2-1 shows the testbed architecture used for outdoor deployment and to evaluate
the combination of mmWave backhaul, mmWave access and EC. The following
sections will explain selected scenarios that we evaluated on the testbeds, with the
corresponding architecture and achieved results.
2.2 Selected scenarios for evaluation
2.2.1 Dynamic Backhaul Reconfiguration
The first evaluated scenario is the dynamic reconfiguration of the mmWave backhaul
links. For this, we prepared the testbed with the architecture shown in Figure 2-2.
Figure 2-1 Application scenario of the testbed
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 10
We already presented a part of the content of this in deliverable D4.2 [D4.2] and in the
paper presented at the BMSB2018 conference [Koslowski2018], but at that point in
time the evaluation was still ongoing and this Deliverable finally contains results that
are more detailed, as published in the Hindawi “Wireless Communications and Mobile
Computing” Journal [Santos2019].
The dynamic backhaul reconfiguration builds a foundation for almost all defined use-
cases and scenarios of the 5G-MiEdge project. With a constantly moving center of
attention in the stadium and a high number of HD cameras generating enormous
amounts of traffic, for the outdoor dynamic crowd that moves throughout a city
requiring high-speed connectivity and for temporarily deployed small cells for events.
With the central orchestration, a dynamic backhaul reconfiguration enables new levels
of flexibility and network optimization. It can quickly detect network failures, switch
to other routes and even repair the topology by re-aligning the links towards other
nodes.
For this, we use the SOCRA (Software-Defined Small Cell Radio Access Network)
architecture, shown in Figure 2-3. It consists of a northbound API, the orchestrator
interface, receiving reconfiguration commands from external applications, the internal
Figure 2-2 Architecture for evaluation of Dynamic Backhaul Reconfiguration
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 11
SDN controller modules, that handle the SDN functionality and the southbound API
to relay commands to the connected nodes.
SOCRA uses a split-plane network design with an out-of-band control channel, the
liquid RAN C-plane.
For evaluation, we used four nodes and placed them throughout the Fraunhofer HHI
lab in Berlin. The nodes were commercially available minicomputers to support the
high throughput rates. We used different brands and generations:
Class 1: Intel NUC Kit D54250WYK (Intel i5 4250U, 16GB RAM, SSD)
Class 2: Intel NUC Kit NUC7i7BNH (Intel Core i7 7567U, 16GB RAM, SSD)
Class 3: Gigabyte BRIX GB-BKi7HA-7500 (Intel Core i7 7500U, 16GB RAM,
SSD)
Nodes N1, N3, N4 are class 2 devices, while node N2 is a Class 1 device. The SDN
controller operates on a dedicated Class 1 device and is connected to the control
network. Both receivers are Class 3 devices and the sender is a Class 1 device.
For orchestration of the tests, we use an additional laptop that is connected to the
control network and uses a REST client application to communicate with the SDN
controller. All devices run on Ubuntu 16.04 with the 4.15.0-34 kernel.
Figure 2-3 SOCRA architecture
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 12
2.2.2 Dynamic Crowd
This section describes demonstration experiments assuming an outdoor dynamic cloud
scenario. Figure 2-4 shows the architecture for evaluation. This testbed has a
combination of the millimeter-wave mesh backhaul in the previous section and the
edge computing function. The orchestrator obtains the movement information of each
user via the microcell base station, and based on that, the SDN controller instructs
migrating the virtual application server to a node (edge) near the user. This allows each
user to receive service with high speed and low latency anywhere in the network. In
this demonstration experiment, they were verified whether the system of migrating
virtual server dynamically operates normally, how much the data rate and delay
improve by comparing the cloud with the edge, and how long it takes to migrate the
virtual server.
2.2.3 Automated Driving
In this section, the testbed structure of the selected EC scenario, i.e., a context-based
dynamic map reception/distribution system with EC, mmWave and SDN technologies
is introduced. Figure 2-5 shows the overall system architecture.
The real-time HD dynamic map of driving environment is a key enabler for safe
automated driving. With dynamic maps measured by 2D or 3D LiDARs, vehicles can
detect and avoid obstacles, navigate and localize moving obstacles in unknown
environments. Moreover, via enabling vehicles to share their local map data during the
cooperative perception [Fukatsu2019] through bi-directional communications, the
driving safety can be improved in a tangible manner.
Figure 2-4 Architecture for evaluation of Dynamic Crowd
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 13
The proposed system aims to provide a safer and more reliable driving environment
for automated vehicles based on EC. In the testbeds, mmWave V2X is utilized for
LiDAR data transmission among EC host and vehicle OBUs, and the SDN technology
is introduced to help manage the mmWave V2X links with a global network view, so
that a better coordination of communication among EC host and vehicle OBUs can be
performed in this network.
2.3 Targeted Features and Details of Scenario-specific
implementation
2.3.1 Dynamic Backhaul Reconfiguration
With the evaluation of dynamic backhaul reconfiguration, we aim to verify the
functionality of the chosen hardware, selected software, implemented algorithms and
the interworking of all those components. We plan to repeat the functional tests to
ensure consistent results. As this feature is one of the main components of both testbeds,
we need to make sure there are no hidden problems.
The features targeted in this evaluation are
Baseline backhaul link reconfiguration
Optimal Steerable mmWave Mesh Backhaul Reconfiguration
Adaptive On/Off Mesh Backhaul Operation
In the Baseline backhaul link reconfiguration, we make sure that the interworking of
all components is accurate, reliable, with a low constant load on the network, and rather
simple configurations.
Afterwards, in the Optimal Steerable mmWave Mesh Backhaul Reconfiguration, we
carefully stress the setup, increase the load on the network, create less ideal conditions
on the channels and evaluate what effects those conditions have on the setup.
Figure 2-5 A SDN based mmWave V2X architecture for safe automated driving
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 14
Thanks to the final feature, Adaptive On/Off Mesh Backhaul Operation, we verify the
power on/off functionality of the nodes and the resulting reconfiguration of the
topology and routes.
All targeted features can be centrally orchestrated by the SDN controller.
2.3.2 Dynamic Crowd
Densification of Small Cell Base Stations (SC-BS) produces massive backhaul traffic
in the core network and employing optical fibers for dense SC-BS backhauling would
result in prohibitively high cost and practical difficulty in implementation. Wireless
backhaul may offer a scalable and cost-effective solution. With GHz bandwidth, the
gigabit data rate is practically achievable, a feature which solves the capacity problem
that exists in lower frequency backhaul systems. In mmWave networks, directional
links are commonly established to compensate for the high path loss and high
directivity paves the way for spatial reuse. Another advantage of mmWave networks
with meshed topology is their capability of flexible traffic routing, which is extremely
suitable for outdoor dynamic crowd, a typical scenario considered where traffic
distribution dynamically varies in both space and time.
On the other hand, as UE nowadays wants to experience services everywhere without
disruption and low latency even when moving, it is desirable that UE-specific
multimedia contents are located as close as possible to the UE via EC technology. The
goal of edge computing is to bring cloud-computing capabilities, including computing
and caching, at the edge of the mobile network, within the Radio Access Network
(RAN), in close proximity to mobile subscribers. Mobile edge applications run as
virtual machines on top of a virtualization infrastructure provided by the mobile edge
host (MEH). Merging mmWave and EC are also essential for other scenarios and
applications, i.e. Olympic Games, multimedia broadcasting services, moving hotspots,
outdoor dynamic crowd and automated driving, as discussed. This paper demonstrates
a content delivery system at the edge over an mmWave meshed backhaul.
The overall testbed systematic architecture is shown in Figure 2-6. It consists of four
pairs of mmWave backhaul links, four pairs of mmWave SC-BSs and UEs, an edge
server (or gateway) which is connected to a global application server, a SDN controller
with an out-of-band control channel, and an optimizer (orchestrator) who orchestrates
the overall network based on UE’s context information e.g. location, application
demands etc. Some tasks of the orchestrator include optimal backhaul route selection,
backhaul link channel allocation and beam selection, prefetching control etc.
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 15
The control protocol for adopting the out-of-band control plane in our work is depicted
in Figure 2-7. Here, the macro base station’s carrier of 2GHz band is adopted as the
frequency band used for the control plane to guarantee a large coverage in order to
centrally manage context information necessary for the C-plane. In addition, in general,
the load of the control plane is sufficiently smaller than that of the U-plane, so narrow
band communications at 2GHz band is sufficient.
2.3.3 Autonomous Driving
The selected scenario incorporates three components:
SDN controller: The central unit, which monitors the network condition and achieves the management logic by creating mmWave links and configuring bi-communication flows of HD map on the components of data plane
OBU: Mobile data plane elements that are controllable by the SDN controller. They are equipped on the automated vehicles, which upload context information and receive control message from the controller to perform cooperative perception
Edge computing host in RSU: Stationary EC data plane elements that are controllable by the SDN controller. They are the infrastructures that are deployed at the intersections or corners (typically on street lamp or buildings). They have a bird view of the traffic condition, and can perform cooperative perception with the SDN OBU
In this selected edge computing scenario, an OBU connects to the backhaul network
of EC hosts in RSUs through mmWave V2X links established by SDN controller
according to its context information. For example, when an incoming OBU arrives at
the edge of a complicated intersection environment, a feasible method to ensure the
driving safety is to get local dynamic maps of those EC hosts in RSUs, which are
Figure 2-6 Systematic architecture of the testbed
Figure 2-7 Out-of-band C-plane for SDN control
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 16
deployed near the intersection. To implement it, the SDN controller will first send
commands to the OBU for an mmWave V2X link with the closest EC host, and then
routes the data flows of maps from different EC hosts to the target OBU. The data
traffic travels across mmWave backhaul and V2X access links. In general, the received
local maps will be merged at the last hop of the transmission to generate a global
dynamic map rather than at each hop for the sake of latency reduction. Therefore, in
this case, the map merging is conducted at the nearest EC hosts. The merged map will
be sent to the OBU via mmWave V2X. Through this process, the mentioned OBU will
obtain a global dynamic map, with which it becomes capable of potential danger
awareness in the out-of-sight area, so that an accident-free driving decision can be
made in advance. This scenario will definitely be an illumination to promote the safety
of today’s automated driving from the perspective of both internal and external system
design.
As for the consideration of the control plane, LTE cellular networks, Wi-Fi hotspots
and WiMAX would be the best choice due to the adequate throughput, wide
communication range and high network densification.
Figure 2-8 shows the composition of functional components that enables our
architecture. Five of them are recognized as key enablers and described below in detail.
Context Manager
A module that enables uploading context information of OBUs to the SDN controller. It helps the controller to maintain a global image of the dynamic network topology. Besides, some safety strategies are made by the controller based on the context of OBUs, e.g. the controller will allow cooperative perception via mmWave V2X when the position of an OBU is close to an EC host.
Link Manager
Figure 2-8 Functional components of proposed architecture
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 17
A module that is triggered by the SDN controller when HD map transmission needs the participation of mmWave V2X links. It creates communication between specific OBUs and EC hosts. It also includes some mechanism that monitor the channel status and perform link repair to ensure the link reliability.
Flow Manager
A module that is triggered by the SDN controller as soon as the link management is done. It guides the map data flow from the source to the destination by sending dynamic OpenFlow tables to specific OBUs and EC hosts in the data plane.
Cooperative perception
A module implemented by the 2D/3D LiDAR with specialized applications for map merging in EC host. It helps an OBU integrate the received local dynamic maps collected from multiple EC hosts to generate a global dynamic map.
Radio Resource Management (RRM)
A module that defines the resource allocation and management logic of our system. It allows the deployment of various RRM algorithms based on the demand for safe automated driving.
3 Deployment of Testbed and Operational Testing
3.1 Dynamic Backhaul Reconfiguration
For the evaluation of this feature, the nodes were deployed in the Fraunhofer HHI lab
in Berlin, as shown in Figure 3-1.
Figure 3-1 Node positions in testbed
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 18
The link distance ranges from 2.4 to 3.8 meters, with unobstructed, direct line of sight
(LOS) links.
As a first step, we performed functional testing of the components, aligned the
interfaces and established links, verified the connectivity and made sure the hardware
was working reliably, also after a reboot or the next day.
The first actual evaluation was the conduction of power measurements. Those are
meant to evaluate the possible power savings when shutting down individual nodes
and reconfiguring the network to keep up with the bandwidth requirements. Power
measurement results are displayed in Table 3-1. This shows an idle power consumption
of approx. 8 Watt, a maximum power consumption of approx. 42 Watt when under full
load and adding another 2-3 Watt when transmitting on both mmWave interfaces for
class 2 nodes.
Power State Class 1 Class 2
Idle without mmWave 4.03 (± 0.22)W 8.01 (± 0.07)W
Idle with mmWave 7.42 (± 0.62)W 11.02 (± 0.13)W
Load without mmWave 17.74 (± 0.60)W 41.70 (± 0.79)W
Load with mmWave 20.12 (± 0.35)W 43.99 (± 1.08)W
Based on this, we believe that significant power saving can be achieved with central
orchestrated power control options for the nodes.
3.2 Dynamic Crowd
Using our developed testbed, an experiment campaign was conducted with preliminary
results to demonstrate the capabilities of the proposed architecture. First, small-scale
indoor test was conducted to confirm functionalities of each component of the testbed.
In the indoor test, since coverage of both U-plane and C-plane is not an issue, reflective
antenna was not used for the backhaul links and Wi-Fi is used as the out-of-band
channel of C-plane. After confirming all components work, the testbed was moved
outside to an about 5000m2 campus area for full-scale outdoor test. In the outdoor test,
due to coverage requirement, reflective antenna was attached to the WiGig USB
dongles for backhaul links and WiMAX was employed as the out-of-band C-plane
channel.
Table 3-1 Power consumption of mmWave nodes
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 19
The full overview of the testbed in indoor environment is shown in Figure 3-2. Again,
for the backhaul links, we use the IEEE 802.11ad WiGig modules, which are produced
by Panasonic Inc., Japan. The modules are accompanied with the driver for Linux
kernel version 4.4.0-36-generic and WiGig supplicant. We configured the module with
MCS 9. The module can operate on the channel 2 and channel 3 among the four
channels of 802.11ad. The maximum value of supported MTU is 7912 bytes. In our
testbed, these modules are connected to a Gigabyte GBBKi7HA- 7500 mini-PC. The
mini-PCs run the Ubuntu 16.04 LTS distribution with Linux kernel version 4.4.0-36-
generic. Each mini-PC has 16GB RAM and 240GB SSD space with a 3.5GHz Intel
processor. For each link, the IEEE 802.11ad module in one side is configured to
operate in the access point (AP) mode, while the other is in the client mode.
We configured a 4 mmWave small cell node testbed topology in a laboratory
environment, where each node is equipped with 2 IEEE 802.11ad dongles. As Fig. 8
depicts, Node 1 (N1) is connected to Node 2 (N2) in channel 2 and to Node 3 (N3) in
channel 3. Node 4 (N4) is connected to N2 in channel 3 and to N3 in channel 2. It is
noted that orthogonal channels were used for adjacent backhaul links in indoor test to
avoid interference while the same channel is used for all backhaul links in outdoor test
due to the difference of link coverage and environment. According to the configured
rules, network traffic between N1 and N4 is forwarded using N2 (path 1, N1-N2-N4)
or N3 (path 2, N1-N3-N4) as next hop. The small cell nodes use Open vSwitch (OVS)
with version 2.7.0 as software switch. We added the IEEE 802.11ad interfaces as OVS
switch ports, and configured two internal interfaces in N1 and N4 to handle network
traffic. In addition, a node running a SDN controller was connected to the mmWave
nodes through a Wi-Fi interface (OF control channel). This controller is based on
OpenDaylight and is responsible for adding and configuring the flow tables of the
managed nodes. We used iperf to generate traffic from N1 to N4, sending 7882 byte
UDP packets during 10 s. Every conducted experiment was repeated 10 times. We used
the ping network utility to measure the round-trip time (RTT) from N4 to N1, as well
as the packet loss/success rates between these two nodes, by sending ICMP echo
packets every 10 ms. Link throughput was measured with tshark in N4, by capturing
network traffic from its mmWave interfaces. Given our testing topology, we confirmed
that the maximum achievable throughput for both 1 hop and 2 hops are about 1.5 Gbps
while the RTTs are 0.835 ms and 1.088 ms for 1 hop and 2 hops respectively.
Figure 3-2 Testbed in an indoor environment
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 20
3.2.1 Outdoor testbed architecture
We deploy 4 nodes in Line-of-Sight environment with a rectangular topology, as
shown in Figure 3-3Error! Reference source not found.. The place is at a university
campus in Tokyo. The nodes are deployed so to have almost the same altitudes for
facilitating beam alignment of backhaul links. In addition, inter-distance between
nodes are about 70m.
The architecture of devices is shown in Figure 3-4. Recapping Sect. II, “Nighthawk
X10 R9000” of NETGEAR were used as APs. “TravelMate P648” of ACER was used
for UE. “GB-BKi7HA-7500” of GIGABYTE for packet transfer and edge computing
were used as mesh node PCs. SFP+ was used as extended M.2-PCIe adapter for
connecting PCs and access points via 10Gbps Ethernet (10GbE) cable. SFX power
supplies were used to activate 10GbE NIC. “WX03” mobile routers of NEC were used
as WiMAX control plane. The GPS receiver for positioning employed “BU-353S4” of
GLOBALSAT.
Figure 3-3 Node deployment in outdoor test
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 21
For orchestrating the experiment, a graphical user interface was designed and
implemented inside the orchestrator PC. It is based on the open source software node-
red, which offers an access to a wide variety of functions, like interacting with the
SDN controller and displaying results in clearly designed and well-structured graphs
functionality without reinventing the wheel. It was configured to issue commands to
the SDN controller, like migrating an EC container from one node to another, and to
the UE, like connecting to a specific AP. Figure 3-5 shows the map view, containing the GPS positions of the available nodes
and the UE. Additionally, the location of the EC container is shown by two colons, at
this time it was running on node 2. The buttons above the map send commands to the
SDN controller to migrate the container to the corresponding node.
Figure 3-4 Hardware architecture in outdoor test
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 22
The GUI also shows the current link performance of the mmWave link of the UE.
Figure 3-6 (a) shows the throughput, which is at roughly 250 Mbps at this time, just
streaming a single HD video. In Figure 3-6 (b) the current latency is displayed, for
the capture duration it showed some slight variations between 0.3 and 0.9 ms and in
average it was around 0.5 ms.
(a) Throughput (b) Latency
3.2.2 Definition of KPIs
The outdoor testbed is evaluated for the KPIs throughput & latency.
For the interface with SDN controller, the orchestrator can request to achieve full
topology information of the network, update each mesh node configuration, decide
migration of EC container to a specific mesh node (via the deploy container button).
Especially, it can even set the MCS of all backhaul links in adaptation to link budget
change due to e.g. rain attenuation. So we evaluate about mmWave backhaul link data
rate and mmWave backhaul link delay.
Assuming that UE is walking toward the upper right EC AP and the UE will start to
play a specific multimedia as it approaches the target AP. If the network can predict
the movement of the UE and migrate the content’s container (Docker) from the
gateway AP to the target AP in advance, the UE can watch the demanded multimedia
Figure 3-5 Map view in the GUI
Figure 3-6 mmWave link measurement on UE
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 23
immediately as it approaches the target AP. Otherwise, the UE needs to play the
multimedia streaming from the Docker located at the gateway AP when it is in the
proximity of the target AP. Of course, for the latter case, the delay is longer since it
includes both backhaul and access delay, while for the former case, the delay is just
incurred by the WiGig access.
3.2.3 Results
To confirm that the mesh network works in even outdoor environment, we conducted
some basic performance test. Table 3-2 shows the backhaul link’s achievable data rate
when the basic MCS 9 is set.
Link Hop number Data rate (Gbps)
Node 1 ⇄ 3 1 1.5
Node 2 ⇄ 4 1 1.5
Node 1 ⇄ 3 ⇄ 4 2 1.5
Node 1 ⟶ 2 1 1.2
Node 2 ⟶ 1 1 1.4
Table 3-3 shows the duration of migrating a sample container with a size of about
5Gbyte over a backhaul link 1→3.
MCS Latency (s)
9 20
5 36
The end-to-end latency between receiver (end user) and EC container was measured
for both cases through the ping command and shown in Table 3-4. The results were
averaged over about 4000 trials.
Type Latency (ms)
w/o EC ≃4
w/ EC ≃0.5
Table 3-2 mmWave backhaul link data rate
Table 3-3 mmWave container migration duration
Table 3-4 End-to-end latency w/ and w/o EC container migration
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 24
3.3 Autonomous Driving
In order to demonstrate the performances of the proposed scenario, i.e., the SDN-based
mmWave V2X network for safe automated driving via coordinating dynamic LiDAR
HD-map communication, we held a proof-of-concept field test in outdoor
environments, where we installed the OBU on a real vehicle, deployed two EC hosts
in RSUs and created a SDN-based network [Li2019].
The Figure 3-7 shows the architecture of the proof of concept, including one OBU,
two RSUs as well as the SDN controller on a remote server. State-of-the-art mmWave
and SDN technologies are applied into this architecture to show their good chemistry.
Moreover, the Figure 3-8 shows the pre-set typical driving scene and path of the OBU,
in which it is going to turn at a corner.
Figure 3-7 PoC system architecture
Figure 3-8 Cooperative perception performed by mmWave V2X
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 25
According to the system design, the SDN-based cooperative perception will be
performed via mmWave V2X along the route of the OBU until it passes the corner.
The whole process is divided into five phases. In the first phase, the OBU just leaves
the starting point. There is no LiDAR data transmission since it is too far from any of
the EC hosts. During the second phase, the OBU enters the coverage of RSU1. The
SDN controller detects its position and triggers the cooperative perception by
connecting RSU1 with RSU2 via mmWave backhaul link, and connecting the OBU
with RSU1 via mmWave V2X. The local dynamic maps of RSU2 and RSU1
simultaneously transmit over the mmWave backhaul and V2X links to the OBU for
integration. Finally, a global dynamic map can be generated at the OBU. Afterwards,
in the third phase, the OBU leaves the coverage of RSU1 so SDN controller
disconnects their link to save power. In the fourth phase, the OBU enters the coverage
of RSU2, but this time, the cooperative perception only occurs between the OBU and
RSU2 via mmWave V2X because the OBU has already passed RSU1, of which the
local dynamic map makes no sense for driving safety. The last phase is quite similar
to the third, i.e., the mmWave V2X link with RSU2 is disconnected by SDN controller
for energy saving.
In order to support such a scenario, the mechanism, which defines the interactive
principles between the data plane and control plane, is described in Figure 3-9.
Through the registration of RSUs, the controller will get knowledge about them (e.g.,
their numbers, positions and mmWave AP information), so that it can prepare the
potential V2X connections for the OBU in advance. As soon as the cooperative
perception is triggered, the context-based flow tables will be assigned by SDN
controller to coordinate the map transmission on EC hosts. Then, the controller will
help the OBU create an mmWave link with the nearest EC host for map reception.
Figure 3-9 Interactive mechanism between data plane and control plane
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 26
Device
Name
Relevant Parameters
Specifications Amount
VLP16 3D
LiDAR
Range: 100 m, Accuracy: +/- 3 cm,
Rotation rate: 5-20 Hz, 16 channels 3
NETGEAR R9000
Tri-Band: 60GHz (4600Mbps), 5GHz
(1733Mbps), 2.4GHz (800Mbps) with 4x
high-performance active antennas
2
mmWave
Antenna Lens model: 5067, Gain: 25.97dBi, Scan
angle: +/- 17deg (Hor.), +/- 4.5deg (Ver.) 2
Panasonic PC
Model: CF-SV OS: Ubuntu 16.04, OVS: 2.8.0, DB
Schema: 7.15.0 2
Acer PC
Model: N15C5 OS: 16.04, ROS: Kinetic, WiGig
Interface: IEEE 802.11, OVS: 2.8.0 1
Vehicle Speed: ≤ 20 km/h (Grey), 0 km/h (Blue) 2
Device
Name
Relevant Parameters
Specifications Amount
WiMAX
NIC
Downlink:120 Mbit/s, Uplink: 60 Mbit/s,
Maximum coverage: 30 miles 4
VAIO PC Model: S13 OS: Ubuntu 16.04, OpenDaylight: Beryllium 1
Table 3-1 Power consumption of mmWave nodes lists the necessary hardware for the
deployment of Data plane. The two 3D LiDARs on RSUs are used to generate HD
maps, and the one on OBU are used to run the Simultaneous Localization and Mapping
(SLAM). Because the NETGEAR router provides 60 GHz mmWave access, we use it
as mmWave AP for the V2X. To constitute the mmWave backhaul between two RSUs,
a pair of mmWave antennas is set up. Two note PCs which support OVS feature allow
the data exchange inside and outside. Another note PC is put on the OBU as it has a
WiGig interface to connect with the mmWave AP at 60 GHz. Figure 3-10 shows their
outdoor deployments. In Table 3-6, besides the PC working as a SDN controller,
WiMAX NICs are used to create a network, which allows the data plane elements
connect with SDN controller to achieve the control logic. Figure 3-11 shows the
hardware deployment of Control plane.
Table 3-5 Hardware for data plane
Table 3-6 Hardware for control plane
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 27
4 Measurements and Evaluation Results
4.1 Dynamic Backhaul Reconfiguration
In this first series of tests, the goal was to test and evaluate the dynamic routing,
configuration and load balancing in the mmWave backhaul network.
4.1.1 Indoor testbed architecture
The used architecture is the same as displayed in Figure 2-2, consisting of 4 mmWave
mesh nodes, with one sender and two connected receivers. There are multiple possible
paths that can be aligned and configured for the traffic between sender and receivers.
When transmitting data from Sender to Receiver 1, the paths are:
1. Sender - N1 - N2 - N4 - Receiver 1
2. Sender - N1 - N3 - N4 - Receiver 1
For data transmission between Sender and Receiver 2, the paths are:
1. Sender - N1 - N2 - Receiver 2
Figure 3-10 Outdoor experimental deployment of hardware: Data plane
Figure 3-11 Outdoor experimental deployment of hardware: Control plane
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 28
2. Sender - N1 - N3 - N4 - Receiver 2
With the installed power control units, the power state of N2 and N3 is centrally
managed to enable power savings and network optimization.
4.1.2 Definition of KPIs
In this series of tests, we aim to evaluate the following KPIs:
1. End-to-end latency
2. End-to-end throughput
3. End-to-end packet loss
4. Reconfiguration delay
5. Power consumption
To guarantee reproducible and accurate results, each test runs for several iterations.
Results
4.1.2.1 Baseline Backhaul Link Reconfiguration
For the baseline evaluation, targeting the reconfiguration delay KPI, we begin with
switching between multiple backhaul configurations.
In this experiment, node N1 is just using a single interface. It is initially aligned to
node N2, like shown in Figure 4-1. With this configuration, Sender and Receiver 1 can
communicate via the path Sender - N1 - N2 - N4 - Receiver 1, shown as a blue dashed
line, and exchange an 800 Mbps UDP flow. With initial measurements, we calculated
the necessary alignment positions and verified them by reading the RSSI on the links
N1 - N2 and N1 - N3.
After transmitting data in this initial configuration for a specified time, the link N1-N2
is disconnected and the interface in N1 is rotated by 70° to face node N3 and establish
a new link, as shown in Figure 4-2.
Figure 4-1 Initial configuration: N1 mmWave interface aligned with N2
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 29
We repeated this experiment for 15 times. Figure 4-3 shows the cumulative distribution
function (CDF) for the different time intervals required to perform the configuration
operations and re-establish end-to-end connectivity between sender and receiver nodes.
These include the mechanical interface alignment from the initial N1-N2 position to
the final N1-N3 position, the internal link configuration by nodes N1 and N3, and the
detection of the newly established link N1-N3 by the SDN controller. To measure the
interface alignment time, we poll the mechanical movement controller periodically
until we receive a reply that indicates the movement has completed, the internal link
configuration is calculated by computing the maximum link configuration time
between AP and STA, ODL detects a new link when it receives an Link Layer
Discovery Protocol (LLDP) packet from one of the corresponding interfaces and lastly
the traffic interruption time is obtained by calculating the maximum time when there
are no packets exchanged between sender and receiver.
The resuts show a constant link alignment time of approx. 3.06 seconds, the internal
link configuration takes an average of 2.88 seconds, due to the used software that
Figure 4-2 Final configuration: N1 mmWave interface aligned with N3
Figure 4-3 CDF for link reconfiguration time
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 30
requires a restart of the wigig_supplicant. After an average of 3.14 seconds, ODL
detects a new link.
These times combined, add up to a total traffic interruption of 6.5 seconds average, but
this is just the baseline configuration with a single interface on node N1, resembling
unexpected outages. Multiple interfaces per node and a redundant mesh topology will
dramatically decrease these timeouts by simply re-routing traffic.
4.1.2.2 Optimal Steerable mmWave Mesh Backhaul Reconfiguration
To implement any wireless backhaul reconfiguration use-cases, it is crucial that the
wireless backhaul can perform these reconfiguration operations with minimal
disruption on existing traffic. This leads to the question: What is the impact of topology
changes and channel assignment on active traffic? To answer this question, we
conducted a number of experiments with two different backhaul configurations C1 and
C2. As a side effect, these experiments validate the long-term functionality of the
hardware and implemented reconfiguration functions in the entire testbed.
Low Traffic Volume: 500 Mbps UDP
The first experiment in this section is using C1 and C2 as shown in Figure 4-4 and
Figure 4-5 and the traffic between sender and each receiver consists of 500 Mbps UDP.
This transition needs to happen when the controller e.g. detects long-term link
blockage on the link N1-N3 and has to re-route the traffic via node N3 instead.
In C1 all nodes are powered up, the sender is connected to N1 and the receivers to N2
and N4. The links N1-N2 and N2-N4 provide paths between sender and the two
receivers.
The second configuration C2 has the same basic topology, but the links N1-N3, N3-
N4 and N4-N2 provide paths between sender and the two receivers.
Figure 4-4 Low Traffic Volume: mmWave backhaul configuration C1
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 31
Figure 4-6 shows the collected measurements of throughput and latency.
At t = 0, the initial setup is performed and configuration C1 is activated, at t = 10, the
measurements start and the throughput rises to the desired 500 Mbps on both paths,
while the latency is rather constant at approximately 1 ms, with a few visible outliers.
To avoid traffic disruptions, the alignment and link configuration of N1-N3 is
performed at t = 30 and t = 40, before updating the flow rules at t = 42. After a very
Figure 4-5 Low Traffic Volume: mmWave backhaul configuration C2
Figure 4-6 Low Traffic Volume: Throughput and Latency
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 32
short timeout of approx. 50 ms, the data transmission between sender and receivers
continues, this is almost invisible in the Figure. While the latency between sender and
receiver R1 stays the same at approx. 1 ms, it rises to approx. 2 ms for receiver r2,
because now there are two intermediate hops. Table 3-1 shows the average and
standard deviation of throughput and latency for both receivers. There is no packet loss
in this experiment.
Events Host Throughput
(Mbps)
RTT (ms) Loss %
Start - Alignment R1 498.65 (±4.76) 1.23 (±0.37) 0.0
R2 498.44 (±4.73) 0.87 (±0.32) 0.0
Alignment - Path 2 config R1 499.82 (±0.34) 1.21 (±0.29) 0.0
R2 499.64 (±0.45) 0.87 (±0.29) 0.0
Path 2 config. - Fwrd. update R1 499.68 (±0.55) 1.25 (±0.37) 0.0
R2 499.55 (±0.46) 0.89 (±0.34) 0.0
Fwrd. update - End R1 499.51 (±0.83) 1.51 (±0.49) 0.0
R2 499.56 (±0.73) 1.70 (±0.57) 0.0
High Traffic Volume: 900 Mbps UDP
The second test is using a similar setup as the first, but the data throughput is increased
from 500 Mbps UDP to 900 Mbps UDP. This resembles situations in real-world
deployments when the network is fully loaded and additional nodes would be powered
on and integrated into the network.
For this, the first configuration C1 is similar to the first test, using the paths:
1. Sender – N1 – N2 – N4 – Receiver 1
2. Sender – N1 – N2 – Receiver 2
But the final configuration C2 changes slightly, as shown in Figure 4-7. Providing the
paths:
1. Sender – N1 – N3 – N4 – Receiver 1
2. Sender – N1 – N2 – Receiver 2
Table 4-1 Low Traffic: Throughput and Latency, average and standard deviation
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 33
In-between C1 and C2, the links N1 – N3 and N3 – N4 are configured and the flow
rules update to re-route the first flow. For this experiment, the results of the bandwidth
of flows F1 and F2, along with the RTT between the server and both receivers, is shown
in Figure 4-8 and detailed in Table 4-2.
In these results, there are can observe interference between the links, causing
significant spikes in the measured latency, as well as packet loss between sender and
receivers.
Figure 4-7 High Traffic Volume: mmWave backhaul configuration C2
Figure 4-8 High Traffic Volume: Throughput and Latency
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 34
Events Host Throughput
(Mbps)
RTT (ms) Loss %
Start - Alignment R1 722.55 (±96.38) 41.59 (±11.29) 19.35
R2 802.52 (±96.63) 42.77 (±8.90) 10.44
Alignment - Path 2 config R1 725.87 (±95.86) 44.89 (±1.54) 19.35
R2 802.94 (±95.70) 44.73 (±1.63) 10.78
Path 2 config. - Fwrd. update R1 740.66 (±97.12) 44.95 (±1.55) 18.04
R2 796.17 (±96.82) 44.75 (±1.65) 11.76
Fwrd. update - End R1 898.33 (±12.55) 1.67 (±3.32) 0.24
R2 896.89 (±5.50) 1.16 (±4.17) 0.46
The situation normalizes when the flow rules for the second configuration C2 are
applied and the traffic for each receiver is using an individual path throughout the mesh
network. Not only does the throughput rise to almost the desired 900 Mbps UDP, the
latency between sender and the receivers also goes down to an average of approx. 1.2
and 1.7 ms.
Interference caused by suboptimal channel selection
To investigate the interference caused by the relatively high number of links for a small
deployment area and the impact of channel selection, this experiment sets all mmWave
links in a path to use the same channel:
N1 – N2, N2 – N4: Channel 2
N1 – N3, N3 – N4: Channel 3
Figure 4-9 and Table 4-3 show the results, and clearly display the high fluctuation in
latency and throughput caused by interference of the active links. Especially when
comparing the results with those of the previous section that actually use the same
positions and links for all nodes, just different channels.
Table 4-2 High Traffic: Throughput and Latency, average and standard deviation
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 35
In addition to the high latency and low throughput, there is a packet loss of up to 75%,
rendering the resulting network almost unusable, at least when aiming for a satisfying
user experience.
Events Host Throughput
(Mbps)
RTT (ms) Loss %
Start - Alignment R1 201.05 (±72.44) 215.75 (±244.96) 73.60
R2 445.69 (±131.86) 80.09 (±73.61) 49.82
Alignment - Path 2
config
R1 189.58 (±72.67) 291.36 (±244.63) 76.50
R2 479.71 (±132.65) 78.14 (±67.66) 45.87
Path 2 config. - Fwrd.
update
R1 213.28 (±87.62) 263.26 (±202.98) 75.34
R2 470.15 (±143.06) 81.75 (±69.47) 48.30
Fwrd. update - End R1 413.24 (±112.18) 139.11 (±129.10) 53.57
R2 890.01 (±27.12) 2.00 (±14.17) 1.33
Only after applying the flow rules for configuration C2, the values improve slightly,
for receiver 2 they even become acceptable, but still the impact of suboptimal channel
allocation is drastic and highlights the need to optimization strategies when deploying
a dense network.
Figure 4-9 Suboptimal channel selection: Throughput and Latency
Table 4-3 Suboptimal channel selection: Throughput and Latency, average and standard
deviation
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 36
4.1.2.3 Adaptive On/Off Mesh Backhaul Operation
To evaluate the power consumption KPI, we used a simplified architecture. There is
just one receiver connected to N4, receiving data from the sender connected to N1. As
shown in Figure 4-10, the first configuration C1 only includes three active nodes N1,
N2, N4, while node N3 is powered off. The path between sender and receiver is Sender
– N1 – N2 – N4 – Receiver and we transmit an 800 Mbps UDP flow.
Figure 4-11 displays the second configuration C2, with node N2 being powered down
and N3 active. This results in the path Sender – N1 – N3 – N4 – Receiver.
For the evaluation, we switch between the two defined configurations and measure the
resulting throughput, latency, loss. To keep the actual interruption to a bare minimum,
the powered down node is powered up first, the interfaces aligned, the links established
before applying the flow rules to switch between the paths.
In Figure 4-12, the measured throughput and latency are displayed, as well as the
timestamps for configuration events.
Figure 4-10 Adaptive On/Off: mmWave backhaul configuration C1
Figure 4-11 Adaptive On/Off: mmWave backhaul configuration C2
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 37
Table 4-4 shows the exact values for throughput and latency, their standard deviation
and the packet loss between sender and receiver.
Events Throughput (Mbps) RTT (ms) Loss %
Start – Alignment 799.12 (±4.90) 1.29 (±0.25) 0.0
Alignment - Path 2 config. 799.75 (±0.64) 1.28 (±0.23) 0.0
Path 2 config. - Fwrd. Update 799.04 (±2.38) 1.29 (±0.36) 0.0
Fwrd. update - Path 1 config. 799.77 (±0.86) 1.27 (±0.24) 0.0
Path 1 config. - Fwrd. Update 799.55 (±0.87) 1.29 (±0.29) 0.0
Fwrd. update - End 799.63 (±0.73) 1.36 (±1.87) 0.0
As shown in these results, the overall performance was consistent and despite the
installation of new flow rules, there was no loss in performance or data.
4.2 Dynamic Crowd
4.2.1 Results on Prefetching
The testbed introduced prefetching algorithm [Nakamura2019] is evaluated via
downloading time of real traffic data. Container migration was realized automatically
by inputting REST API command to SDN controller at the set time, which is based on
the scheduler registered beforehand by UE. The scheduler was determined based on
user moving speed, throughput of access and backhaul, container start-up time, etc.
which are shown in Table 4-5. Each MEH and UE input REST API commands to SDN
controller to implement the prefetching algorithm based on the scheduler.
Parameter Value
Figure 4-12 Adaptive On/Off: Throughput and Latency
Table 4-4 Adaptive On/Off: Throughput and Latency, average and standard deviation
Table 4-5 Simulation parameters for schedule creation
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 38
Throughput of Access [bps] 1.0 × 109
Throughput of Backhaul [bps] 7.0 × 108
Data size of image [Byte] 2.0 × 109
User speed [m/s] 0.5
Inter-cell distance [m] 25
Container start-up time [s] 40
Waiting time for watching video [s] 20
The schedule of prefetching algorithm and experiment environments are shown in
Figure 4-13 and Figure 4-14 respectively. At first, UE stays at Node1 and connects to
the nearby AP using WiGig. After UE finished downloading 4K-Videos, UE waits a
few seconds for watching the movie. During then, each node starts container migration,
which includes a process of stopping container, transferring container image among
the MEHs and starting up container at another MEH. After watching the video, UE
starts moving from the source node to the destination node. The testbed was designed
to finish container migration just before UE arrives at the next node.
The measurement result is shown in Figure 4-15. In case of EC migration utilizing
prefetching algorithm, it was confirmed that the downloading time of 4K-Video did
not change even if the user moves to the other node. On the other hand, download time
increased as the number of hops increased when the content server is fixedly deployed
on Node 1, in other words, the without edge computing (and migration) case.
Especially, download time increased when user arrived at Node3 for this case. Packet
error rate increased due to multi-hop relay with the increase of re-transmission requests
might be the reason for the worsen phenomenon of w/o prefetching. Detailed
investigation is considered as our future works.
Figure 4-13 Schedule of prefetching algorithm based on the parameters
Figure 4-14 SC-BSs deployment and user movement
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 39
4.2.2 Results on Route Multiplexing
In addition, the testbed is evaluated by backhaul route multiplexing. When intensive
traffic distribution is happened, the architecture applies several backhaul resources to
improve access throughput. By using SDN control, the architecture can realize
adaptive allocation of backhaul resources to follow dynamic variation of traffic
distribution.
In this experiment, EC deploys at node 1 and node 3 in according to Figure 4-16.
Maximum backhaul throughput is about 1 Gbps, and maximum access throughput is
about 2 Gbps, so there is bottleneck in case there is only one backhaul route available
like the left figure. There will be no bottleneck in case there are several backhaul routes
like the right figure. Thus, the effectiveness of mmWave access can be confirmed.
Figure 4-15 Downloading time of 2GB movie data with FTP
Figure 4-16 Architecture of backhaul route-multiplexing
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 40
The measurement result is shown in Table 4-6. In case of “w/o Route-Multiplexing”,
route1 (EC – Node1 – Node3 – Node4 – UE) and route2 (EC – Node3 – Node4 – UE)
use the same backhaul route (Node3 – Node4). As mentioned above, there is bottleneck
in this case. The total throughput is 811 Mbps, which is not the maximum access
throughput but approximately the backhaul link’s throughput. In case of “w/ Route-
Multiplexing”, route1 (EC – Node1 – Node2 – Node4 – UE) and route2 (EC – Node3
– Node4 – UE) use different backhaul routes, therefore there are no bottleneck. The
total throughput achieves 1.56 Gbps, which is much higher than single backhaul
throughput and nearly the maximum access throughput. Comparing these two
scenarios, the total throughput of “w/ Route-Multiplexing” becomes 1.9 times as
higher as that of “w/o Route-Multiplexing”. Therefore, we confirmed the effectiveness
of backhaul route multiplexing using SDN control.
Method Route Throughput Total throughput
w/o Route-
Multiplexing
EC – N1 – N3 – N4 – UE 427 Mbps 811 Mbps
EC – N3 – N4 – UE 384 Mbps
w/ Route-
Multiplexing
EC – N1 – N2 – N4 – UE 816 Mbps 1.56 Gbps
EC – N3 – N4 – UE 743 Mbps
4.2.3 Results on Integration with Internet and Wi-Fi
In the next section, by changing the access method and the number of hops, we again
check the effectiveness of edge computing in mmWave cellular network. Figure 4-17
shows an illustration of the measurement patterns. Three among the four available
nodes, one container having iperf server, and one UE are used in this testbed. In 0-hop
pattern, the container is set in node4, which is the nearest position to UE. In 2-hop
pattern, the container is set in node1 and UE connects to the server via one access link
and two consecutive backhaul links. In the both patterns, two access methods (Wi-Fi,
WiGig) are used. In multi-hop pattern, there is no container in this setting. Node1 is
connected to the Internet and UE connects to an iperf server in Germany via multi-hop
links.
Figure 4-18 shows the results of throughput measurement. In the case of using Wi-Fi
access, throughputs are about 600Mbps with no hops and two hops. These results mean
even if EC is used, Wi-Fi become a bottleneck and it is impossible to achieve
throughput higher than 600 Mbps. On the other hand, in the case of using WiGig access
with two hops and multi-hops, throughputs attain about 900 Mbps and 30 Mbps
respectively. These mean even if WiGig is used, backhaul link becomes a bottleneck
and it is impossible to demonstrate original performance of WiGig. However, by
fetching server in GW node (node1), backhaul bottleneck issue can be resolved. Finally,
in the case of using WiGig access with no intermediate hops, the throughput achieves
the fastest speed of about 1.3 Gbps. This means WiGig can demonstrate its full
potential by setting server in the closest node to UE.
Table 4-6 Throughput of route-multiplexing with iperf3
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 41
4.2.4 Results on Multiple UEs
Then, we demonstrate the benefits of deploying multiple containers in environments
with multiple UEs. Figure 4-19 shows measurement configuration of multi-UE and
multi-container scenario. In addition, Table 4-7 shows throughput measurement results.
mmWave WiGig is used as access link and two UEs connect to node1 and node2
respectively. In case of single container, the container is set in node1. UE1
communicates to the server directly via WiGig access. UE2 simultaneously
communicates to the same server via WiGig access and one backhaul link. Normally
in single-UE scenario, UE1 has throughput of about 1.3 Gbps and UE2 has throughput
of about 900 Mbps. However, in this multi-UE experiment, the throughputs of the UEs
Figure 4-17 Measurement patterns
Figure 4-18 Results of throughput measurement with iperf
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 42
are reduced to about 900 Mbps and 350 Mbps respectively when two UEs
simultaneously access to the server through different routes. On the other hand, in
multi-container scenario, two containers are deployed in node1 and node2 in a
distributed manner. Two UEs connect to the containers directly via WiGig access.
Thereby, the throughputs can achieve higher value compared to the multi-UE scenario
since the effect of backhaul bottleneck and routing degradation can be resolved.
Container pattern UE index Throughput [Gbps]
Single 1 0.90
2 0.35
Multi 1 1.22
2 1.15
4.2.5 Results on blocking mmWave access
Blocking effect on backhaul link can be avoided by reconstructing route or migrating
EC server to edge node in our network configuration. A situation when a human body
blocks mmWave access link is examined in this part. Netgear R9000 Nighthawk
wireless router we use in this examination has a function of beamforming so that
optimal path can be always formed. Figure 4-20 shows measurement environment in
this blocking situation. There are a node with mmWave access point (AP), an UE and
a human by a concrete wall. The UE and the mmWave AP are set at a height of 1m,
the distance between them is 5m. The wall is set in parallel and 1m away from them.
In this environment, the UE downloads data from the mmWave node.
Figure 4-19 Measurement configuration of multi UEs and multi containers
Table 4-7 Throughputs of multi UEs and multi containers
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 43
Figure 4-21 shows the fluctuation of throughput. At first, the throughput has high value
because direct beam path is formed. A few seconds later, the human blocks the path
and the throughput drops sharply. However, by re-adjusting the beam according to BTI
(beam transmission interval), a path using wall reflection is selected, and throughput
returns to the high value. Namely, from this result, blocking mmWave access is not an
issue in environment with walls.
Figure 4-20 Measurement environment in blocking situation
Figure 4-21 Fluctuation of throughput in blocking situation
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 44
4.3 Autonomous Driving
We will give an illustration about how safety this system can provide from the GUI,
which presents the received and merged dynamic HD maps on the EC hosts. As shown
in Figure 4-22 (a), in Phase 1, only OBU map (green points) appears, because the
cooperative perception has not been performed since it is out of the coverage of
mmWave V2X AP. Nevertheless, after the OBU enters the coverage of RSU1, as
shown in Figure 4-22 (b), relying on the cooperative perception with RSU1 local map
(red points) and RSU2 local map (yellow points), its vision is extended to the
surrounding area of the corner. At the same time, an incoming vehicle, which is pointed
by an arrow in Figure 4-22 (b), in the blind spot is detected so that the OBU can take
measures to avoid collision in advance. From Phase 2 in Figure 4-22 (b) to Phase 4 in
Figure 4-22 (c), due to the dynamic topology, the corner environment may have
changed. So cooperative perception performed again with RSU2 local map (yellow
points) to update the global maps.
The end-to-end throughput between the OBU and RSU2 was measured because this
parameter features remote and huge data transmission at the same time. We test the
performances of 2.4 GHz and 60 GHz as candidates of V2X respectively and compare
their performance. As Fig.21 shows, in Phase 1, 3, 5, the V2X communication was not
established because the OBU was moving out of the coverage. When the cooperative
perception was first triggered in Phase 2, the 60 GHz, i.e., mmWave V2X can quickly
provide a data rate of near 1 Gbps while the achievable data rate of 2.4 GHz V2X was
only 90-100 Mbps. In Phase 4, cooperative perception was performed again as SDN
controller detected the OBU was within the range of RSU2. Although there was no
map transmission through the backhaul links because the RSU1 did not engage, the
throughput was almost the same as Phase 1.
a) Phase 1 b) Phase 2 c) Phase 4
Figure 4-22 GUI of received dynamic LiDAR maps
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 45
4.4 Integration of 5G!Pagoda and 5G-MiEdge testbeds
5G!Pagoda is a European/Japanese research project under the H2020 call, like 5G-
MiEdge is. Throughout the projects, we had continued discussions between the
members of both consortiums and planned to investigate possible benefits when
combining our SDN orchestrated mmWave edge computing with their SDN
orchestrated network function virtualization (NFV) [D1.4].
To summarize our discussions, we will show some measurements of the virtual
integration of 5G-MiEdge and 5G!Pagoda networks in Figure 4-24. Wired SDN
constructed by 5G!Pagoda decreases the latency from 400ms to 175ms, and wireless
SDN constructed by 5G-MiEdge decreases it from 4ms to 0.5ms (in a local domestic
network) as shown in Tab. 4-8. Therefore, we confirmed the effectiveness of
wired/wireless SDN and edge computing.
Figure 4-23 comparison of the network performance by using V2X of different frequencies
(2.4/60 GHz) during the PoC
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 46
Method Latency [ms]
Access to foreign server via general network 400~500
Access to foreign server via 5G!Pagoda network 175
Access to local domestic server w/o 5G-MiEdge’s EC 4
Access to local domestic server w/ 5G-MiEdge’s EC 0.5 [Table 3-4]
Figure 4-24 Integration test of 5G-MiEdge and 5G!Pagoda networks
Table 4-8 Comparison of 5G-MiEdge and 5G!Pagoda integration network by latency
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 47
5 Final Demo at EuCNC
As this is the final year of the project, we went to the EuCNC2019 in Valencia, Spain
to exhibit our testbed hardware in a final demo, displayed in Figure 5-1.
Due to the space and transportation constraints, we went for a slightly reduced
architecture, as shown in Figure 5-2.
This final demo is using three mmWave small cells, as shown in Figure 5-3, equipped
with two movable mmWave backhaul interfaces in the top compartments, a
computation device and with nodes N1 and N2 being also equipped with multi-RAT
mmWave/sub 6 GHz access in the lower compartment.
Figure 5-1 Final Demo at EuCNC2019
Figure 5-2 Architecture at EuCNC2019
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 48
The UE is represented by a laptop that connects to the APs via multi-RAT
mmWave/sub 6 GHz and requests data from the EC hosts.
We implemented parts of all previous evaluations in this final demonstration to
summarize our efforts and give an overview of the implemented functionality.
For this demo, we chose a static backhaul topology with three links:
1. N1 – N2
2. N1 – N3
3. N2 – N3
The links were always aligned and connected. The SDN controller, however, was able
to re-route the traffic between these links, depending on the link quality, indicating
obstructions or completely blocked links.
The laptop was constantly requesting content from an edge computing application
running on node N3. In the case of mmWave access link blockage, the multi-RAT
functionality seamlessly switched to the sub 6 GHz access link, resulting in slightly
higher latency and lower throughput, but still ensuring connectivity to the service. For
link failures longer than 3 seconds, the SDN controller orchestrated the laptop to
switch to an AP running on a different node and restore low-latency connectivity at
full speed.
This resulted in four possible routes that the SDN controller automatically selected:
1. UE – N1 – N3
2. UE – N1 – N2 – N3
3. UE – N2 – N3
4. UE – N2 – N1 – N3
Figure 5-3 5G-MiEdge node used at EuCNC 2019
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 49
After running for four days, we were quite satisfied with the performance and
reliability. Despite blocked links in backhaul and access, the connectivity between UE
and EC application was quickly restored, providing a high end-to-end link reliability
and quality of service.
The conference was one of this year highlights and we had many interesting
discussions with visitors and researchers from other projects!
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 50
6 Summary
This deliverable presented the outcome of Task 4.2 ‘Development of common/joint
5G MiEdge Testbed’ and Task 4.3 ‘: Field trials toward Tokyo Olympic 2020’. It is the
final document in WP4 and marks the end of this project.
We conducted successful indoor and outdoor evaluations in our testbeds in Berlin and
Tokyo, not only showing that our developed technologies are ready for events like the
Tokyo 2020 Olympics, but also targeting general 5G deployments and current verticals
as autonomous driving.
The results and evaluation of our developed concepts were based on solid research
from other tasks in WP4, as well as the foundation in WP1, WP2 and WP3.
For the future, we are looking forward to extend our developed prototypes and
achieved results with new concepts, continue our partnership with members of this
consortium and research upcoming technologies!
Deliverable Horizon2020 EUJ-01-2016 723171 5G-MiEdge D4.3
Date: July 2019
Public Deliverable
5G-MiEdge Page 51
7 References
[D4.2] 5G-MiEdge – Deliverable D4.2 “5G MiEdge test-bed integrating mmWave access,
liquid RAN C-plane, and user/application centric orchestration”
[Koslowski2018] Konstantin Koslowski et al “SDN Orchestration to Optimize Meshed Millimeter-Wave
Backhaul Networks for MEC-enhanced eMBB Use Cases”, BMSB2018
[Santos2019] Ricardo Santos et al, “mmWave Backhaul Testbed Configurability Using
Software-Defined Networking”, Hindawi Wireless Communications and Mobile
Computing, Volume 2019, Article ID 8342167
[Tran2019] Gia Khanh Tran et al, “Outdoor Experiment of mmWave Meshed Backhaul for Realtime
Edge Content Delivery,” IEEE WCNC2019, Apr. 2019.
[Fukatsu2019] R. Fukatsu et al, “Millimeter-Wave V2V Communications with Cooperative Perception
for Automated Driving,” IEEE VTC-Spring, May 2019.
[Nakamura2019] M. Nakamura et al, “Performance Evaluation of Prefetching Algorithm for Real-time
Edge Content Delivery in 5G System,” IEEE VTC-Fall, Sep. 2019.
[Li2019] Z. Li et al, “Proof-of-Concept of a SDN Based mmWave V2X Network for Safe
Automated Driving,” IEEE Globecom2019, Dec. 2019.
[D1.4] “Final report on joint EU/JP vision, business models and eco-system impact,” 5G-
MiEdge Deliverables, Jul. 2019.