2021 30th Wireless and Optical Communications Conference (WOCC)

5G Edge Computing Experiments with Intelligent

Resource Allocation for Multi-Application Video
Analytics
Tzu-Hsuan Chao, Jian-Han Wu, Yao Chiang, Hung-Yu Wei
2021 30th Wireless and Optical Communications Conference (WOCC) | 978-1-6654-2772-2/21/$31.00 ©2021 IEEE | DOI: 10.1109/WOCC53213.2021.9603242

Dept. of Electrical Engineering

National Taiwan University
Taipei, Taiwan

Abstract—The fifth-generation mobile network is characterized an important milestone in the 5G era. Furthermore, live video
as the edge of wireless connectivity for all intelligent automation. analytics can be a common application in the MEC system
Technically, the services’ requirements for Quality of Service because of low-latency demand. For example, surveillance
(QoS) have become more strict on latency and throughput. As
a result, the concept of Mobile Edge Computing (MEC) has cameras play an essential role in maintaining public safety. In
become promising. By placing servers close to the user-equipment addition, the widespread development of traffic monitors, self-
(UE), the paradigm enables much lower data transmission time driving, Virtual Reality (VR), and Augmented Reality (AR)
compared to the cloud-based scenario. With this advantage, MEC related to video analytics. Object detection and recognition
reaches the requirements of low-latency. Moreover, recognition apply in many areas of computer vision, including image
and detection technology can be thus implemented in several live
video analytics scenarios. However, due to the limited physical retrieval, security, surveillance, automated vehicle systems,
size on the edge server, resource allocation becomes a crucial and machine inspection.
issue. In this paper, we proposed a Resource Management method However, limited computing resources of edge servers may
with Multiple Applications in Edge architecture (RMMAE) to not afford multiple complexity and huge DL models. In brief,
intelligently reallocate computing tasks in the heterogeneous it is a trade-off problem between low latency and high quality.
network. We design an algorithm to allocate computing resources
to applications such as facial detection, object detection and Delivering all data to the cloud server and performing high-
pose estimation in our Edge testbed, and we prove impressive quality video analytics come with exceptionally high trans-
improvement and performance on our testbed with multiple mission costs. If we only process the data to the edge server,
applications. the server near users, the computing power may constrain the
Index Terms—Edge Computing, Resource Allocation, Live result. So, we choose edge computing of cloud computing
Video Analytics
on demand. Some works proposed the idea that determines
whether the service is quality-oriented or speed-oriented by
I. I NTRODUCTION
scoring users’ needs, such as [2] and [3]. The distributed
With the rapid rise of the fifth-generation mobile commu- computing for the edge server and the cloud server is another
nications era, artificial intelligence and Deep Learning (DL), choice. The concept is similar to federated learning [4]. The
the efficiency and speed of network transmission and powerful authors applied a decentralized approach to make the system
computing capabilities are becoming increasingly important. more efficient. Both [5] and [6] deployed the work through
Therefore, cloud computing is also popular because of its distributed deep learning models deployed on the MEC nodes
sufficient data storage and excellent computing power without for better performance. We can divide one task into sub-tasks
direct active management by the user. Nowadays, several and distribute them to an edge server and a cloud server
providers have paid service of the cloud server, such as for computing in an Edge-Cloud system. On the other hand,
Amazon Web Services (AWS), Google Cloud Platform (GCP), resource allocation is also an essential issue for the edge-cloud
Microsoft Azure, etc. Users may choose suitable services system. Hence the restriction of the computing resource on the
to solve the problems. Nevertheless, the weakness of cloud edge server, we need to allocate limited resources effectively
servers that the transmission delay between end-users and and use them where it is most need. [7] proposed a block-
cloud servers may keep the cost increasing considerably. Video chain-based video streaming system, and they formulate the
analytics services in real-time are also harsh constraints on problem of optimizing the offload scheduling to optimize. [8]
network bandwidth and congestion. Therefore, the purpose provided a promoted QoS method to offload tasks.
of our works is to provide a high-quality and low-latency The DICE-IoT system in [9] is our previous work. It
computing system. Base on the ETSI Mobile Edge Computing provides a facial recognition app for video analytics and for-
(MEC) in 5G white paper [1], we realize the edge computing is mulates the management to incentivize cooperative computing
provision between the Edge and the Cloud. Inspired by this
The authors are grateful for the funding support from Ministry of Science
and Technology (MOST) of Taiwan under grant 109-2221-E-002-148-MY2 work, we will continue the concept of resource allocation
and 109-2218-E-002-018-. and implement more applications on our testbed to deal with

Authorized licensed use limited to: VTU Consortium. Downloaded on October 22,2024 at 05:07:11 UTC from IEEE Xplore. Restrictions apply.
978-1-6654-2772-2/21/$31.00 ©2021 IEEE 80
 2021 30th Wireless and Optical Communications Conference (WOCC)

TABLE I
PARAMETERS OF CONFIGURATION

Notation Definition
f rφ Frame sampling rate (in fps)
wφ , hφ Frame {width, height} (in pixels)
rsφ Video resolution rsφ = wφ × hφ
bφ
edge Data size uploaded
from the device to the edge node per frame (in MB)
bφ
cloud Data size output
from the edge node to the cloud per frame (in MB)
mφ
edge GPU memory usage at the edge node (in MB)
mφcloud GPU memory usage at the cloud (in MB)
Fig. 1. Hierarchical architecture of two-tier Edge-Cloud in our testbed uφ
edge GPU utilization at the edge node (in [0, 1])
uφ
cloud GPU utilization at the cloud (in [0, 1])
cφ
edge Energy consumption at the edge node (in Watts)
more complex problems. Furthermore, we consider a model- cφ Energy consumption at the cloud (in Watts)
cloud
level pipeline composition that can decompose the model on pφ Per frame processing time at the edge node (in ms)
edge
the different servers to cooperate for one application. The pφ Per frame processing time at the cloud (in ms)
cloud
architecture of our MEC system is a two-tiered edge cloud
computing with real-time video analytics. On top of that,
we provide an improved method RMMAE for resource load while the cloud server is assumed to have unlimited resource
balancing and implement the complete MEC system in the budget.
actual environment. We considered three types of end devices: mobile phones
The main contribution of our work is as follows: connected to a 4G/5G base station, those connected to a WiFi
• We reduce the transmitted latency by allocating resources access point, and PCs connected with the Ethernet; each of
appropriately according to the system’s current situation. them has a camera. The solid line is the data flow, and the
• This is the Edge system with implied reasonable algo- dashed line means is the control flow. The camera records
rithms to discuss and realize the situation where multiple live video and streaming it to the MEC server for real-time
applications are in the system at the same time. analytics. The control nodes and the orchestrator in the MEC
• We distribute the resources by the detection accuracy server will then allocate computing resources and forward the
request, current video transmission quality and GPU streaming request to either an appropriate MEC node or the
status of edge server. cloud server, based on the end users’ demand and the GPU
status of the edge server. Generally, the edge server aims to
II. P ROPOSED S CHEME D ESIGN provide the service with high quality and low latency.
In this section, we introduce the three-layered architecture of
the system, including the Cloud Layer, the MEC Layer and the B. System Model and Configurations
User Application Layer. The interaction between these layers Our video programs run in both Cloud and MEC Layers,
allows better allocation and improves the system performance. but the resource management is only considered in MEC due
Also, we design the data flow and the control flow with our to the infinite resource assumption in Cloud. The devices in
management mechanism. the User Application Layer capture video and streaming it to
the MEC node for real-time video analytics, such as object
A. Edge-Cloud Cooperative Computing Architecture detection, facial detection, and pose estimation. Suppose that
Edge and Fog Management and Orchestration is essential each device has a specific latency requirement Lreq and a
in the MEC Layer, which plays a critical role in coordination quality requirement, Qreq . We also define configuration φ as
between components. According to hierarchical management, a specific decomposable pipeline and Φ as the collection of
we can master and integrate all the MEC nodes easily and all configurations.
orderly. Figure 1 shows the architecture of the Edge system We also consider the resolution and the video frame sam-
in our testbed. First, the Edge server includes several MEC pling rate for all devices and the GPU resources as the
nodes and the Edge Orchestrator; then, the Edge Orchestrator computing power on the cloud server and the edge server for
will manage and coordinate MEC nodes to work in the MEC each configuration φ. The partial offloading indicator ϕφ is a
server. As for a single MEC node, it consists of a control binary indicator and is defined as follows: If ϕφ = 1 means
node and a compute node. The control node may receive the the video is uploaded to Cloud Layer for serving. Otherwise,
control signals sent by the orchestrator. It also collects and MEC Layer processes the task.
sends the information about GPU status to the orchestrator To sum up, we denote the complete configuration set
periodically. The compute node in the MEC node provides Φφ (rsφ , f rφ , ϕφ ), which consists of all available configura-
low latency computing, but with limited resource constraint tions for the proposed DICE scheme. On the other hand, we

Authorized licensed use limited to: VTU Consortium. Downloaded on October 22,2024 at 05:07:11 UTC from IEEE Xplore. Restrictions apply.
81
 2021 30th Wireless and Optical Communications Conference (WOCC)

selections of MEC nodes. The binaryn variables x stando for

φ
user association decision, where x = xij , ∀ i, j, φ . The i
and j are the number of devices and MEC nodes, respectively.
xφij = 1 means the device i is associated with the MEC node
j with configuration φ for video analytics; otherwise xφij = 0.
We analogize the network to a directed graph G (M, L),
where M means the set of MEC nodes and L indicates the
communication between devices and the servers.

A. Social Welfare derivation

max W (x, n) = Umec + Ucld

 X X φ

 xij ≤ 1 ∀i
 c φ∈Φc


 j∈M i j

Fig. 2. Workflow of periodic resource allocation control signals  X X φ

xij = 0 ∀i
s.t. C1 : c φ∈Φ

 j∈M\M i
 X X
xφij = 0 ∀i

also denote the configuration set, which
 only consists of the




MEC stratum configuration Φmec = φ : ϕφ = 0, ∀φ ∈ ΦU j∈M φ∈Φ\Φcj

We may select the configuration according to the resource  φ φ
status of the MEC node, latency requirement, and accuracy nj ≤ Nj ∀j, φ

 X
nφj ≤ xφij ∀j, φ


requirement. Table 1 organize the parameters list which we 



introduce in our problem formulations. 
 i∈N
C2 :
X φ φ
C. Work flow for control signals  nj uj ≤ uavl j ∀j


 φ∈Φ
Figure 2 presents the periodic control signals between the 
X
nφj mφj ≤ mavl ∀j


devices, the MEC nodes, and the cloud server. The period 

 j
φ∈Φ
starts with communication between the servers and the user
device. After that, the user QoS report and the server state are X nφj X
sent to the orchestrator. Next, the orchestrator will apply our C3 : f rφ xφij ≤ =⇒ f rφ tφj xφij ≤ nφj , ∀j, φ
i∈N tφj i∈N
RAMME method to balance the load between the edge server  
and cloud server, and to select proper MEC nodes and the C4 : xφij Lreq
i − Lφij ≥ 0 ∀i, j, φ.
corresponding computing scheme. In addition, it will configure X X
data forwarding rules for data managers. Therefore, the system C5 : xφij f rφ bφ ≤ rij , ∀lij ∈ L
j∈M φ∈Φ
updates data flow and the connection between MEC nodes,  X X
cloud servers, and user devices. In conclusion, the orchestrator  xφ f rφ bφ ≤ rij , ∀lij ∈ L
 l ∈L φ∈Φ ij


will manage the edge server and also communicate with ij
C6 :
the UE apps while the system will update the computing
X X φ

 xij f rφ bφcld ≤ rj,cld , ∀j
configurations and offloading scheme in a more suitable way.


i∈N φ∈Φ
We can then control and adjust the system’s status dynamically n o
to make services more advanced and efficient. xφij ∈ {0, 1} , nφj ∈ 0, 1, · · · , Njφ ∀i, j, φ
(1)
III. A LGORITHM AND P ROBLEM F ORMULATION where the Umec and Ucld are the utility of MEC Layer and
We apply the intelligent algorithm RAMME for joint utility of Cloud Layer. Constraints C1 and C2 represent the
latency- and accuracy-aware live video analytic services with user association between UEs and MEC nodes, and they are
multiple applications in our actual testbed. Referring to the bounded by the device requirements and GPU status. C3
previous work [9], this paper not only considers several types considers the frame rate as the constraint to maintain stability
of communication connected to our edge server but imple- of service. Since the latency requirement is asked by the device
ments kinds of applications for users to choose. The problem i, C4 shows the latency constraint. Furthermore, the data rate
we’ve solved is to balance the QoS and the service latency. of each link needs to satisfy its transmission capacity by C5
Therefore, we develop a scoring mechanism considering both and C6.
social welfare to measure the revenue, and the cost on the Edge
and Cloud, to judge the performance of the offloading scheme. B. Utility of Cloud layer
The expected socialnwelfare is o
maximized by selecting optimal Since the MEC stratum pays π to compensate the per unit
n, x, where n = nφj ∀j, φ represents the configuration effort to the cloud for the optimal computing provision, the

Authorized licensed use limited to: VTU Consortium. Downloaded on October 22,2024 at 05:07:11 UTC from IEEE Xplore. Restrictions apply.
82
 2021 30th Wireless and Optical Communications Conference (WOCC)

utility function of the cloud may like: A. Deploy MEC system in our Testbed
1) Testbed Environment: Chunghwa Telecom (CHT) has set
Ucld = Rcld − Ccld (2) up a 5G campus network and MEC server. The edge server
of our experimental testbed is point-to-point connected with
Ccld means the operation cost of the cloud per second: Chunghwa Telecom’s MEC server to develop various scenarios
X X X φ for the main architecture of the system.
Ccld = γcld xij priceφcld tφcld (3) MEC server: The server consists of a control node and
i∈N j∈M φ∈Φ a compute node. The compute node is equipped with an
individual GeForce RTX 3080 GPU instance. As for external
The γcld is a constant that denote the cost of converting
connections, there is a CHT SDN switch link between the
power consumption at the cloud server. We regard the total
MEC server and three kinds of UEs.
GPU utilization as the effort per second of the cloud.
Cloud server: Elastic Compute Cloud (EC2) of Amazon
X X X φ φ φ φ Web Service (AWS) is a cloud-server rental service of Ama-
ecld = xij tcld ucld f r (4)
zon. We rent Amazon EC2 G4 Instances and its type is
i∈N j∈M φ∈Φ
G4dn.xlarge equipped with NVIDIA T4 GPUs and custom
And the revenue of the cloud present as: Intel Cascade Lake CPUs.
User Devices: We have three scenarios to simulate the
Rcld = πecld (5) user devices: PCs connected to the Ethernet to surf the
C. Utility of MEC layer Internet, laptops to Wi-Fi, and mobile phones to 4G/5G mobile
networks. PCs and laptops connect to the MEC server via the
The expected utility function of the MEC stratum is: campus internet, while phones connecting via the CHT fiber
Umec = Rmec − Cmec − Pmec (6) optic link from the base station to the MEC server.
Applications: Our test platform has three types of ap-
plications: facial detection, object detection, and pose es-
And Pmec = Rcld represents the total service fee. timation. In our previous work [9], we provide the facial
The operation cost per second of the MEC stratum is similar detection application. This paper implements object detection
to the Cloud stratum: and pose estimation with YOLOv4, YOLOv4-tiny AlphaPose,
X X X φ φ φ
Cmec = γmec xij pj f r (7) and OpenPose. Also, we combine different kinds of video
i∈N j∈M φ∈Φ
analytics models above for a decomposable inference pipeline
and deploy with three offloading types: Edge only, Cloud only,
The γmec is a constant that denote the cost of converting power and Edge-Cloud.
consumption at the MEC node. We assume the service fee is 2) Parameter Measurement in Testbed: Taking the actual
proportional to the user satisfaction, we have the revenue as situation like the scene, we use the mobile phone with CHT
follows: X X X φ 5G network to capture the video stream as input and adjust
xij P q φ


Rmec = (8) the relevant camera and transmission information such as
i∈N j∈M φ∈Φ frame rate and video resolutions to measure various situations.
Both the edge server and cloud server have individual GPU
instances as in the previous description. We list 70 kinds of
And according to equation (3) and (7), the social welfare configurations with the different model schemes, and three
can be derived as: offloading types: Edge Only means the inference pipeline ran
X X X φ at the MEC layer entirely, Cloud Only means the inference

xij P q φ − Ωφj


W (a, x, n) = W (x) =
i∈N j∈M φ∈Φ
pipeline ran at the Cloud layer entirely, and Edge-Cloud, the
(9) inference pipeline is decomposed to run at the MEC layer and
Cloud layer.
The total operation cost 3) Testing Scheme: We simulate the user requests and
  implement the RMMAE system in the actual testbed based
Ωφj = Υmec pφj + Υcld pφcld f rφ , ∀j, φ (10) on the configurations. The simulation parameters are as fol-
lows: the frame rate as 30 fps, the video resolutions are
uniform distribution in 1080p, 720p, and 480p, the accuracy
requirement is the uniform distribution between 0.5 to 0.8, the
IV. E XPERIMENTAL R ESULTS latency requirement is the uniform distribution between 200
This chapter set up an implementation platform and applied to 1200(ms) and the application selection also random in three
the RMMAE algorithm. There are multiple video analytic choices. Further, we compare six schemes for experiment:
applications on a real testbed to select and prove our MEC (1) Proposed-C: Completely framework with enable pipeline-
system’s better performance by the experimental results. sharing.

Authorized licensed use limited to: VTU Consortium. Downloaded on October 22,2024 at 05:07:11 UTC from IEEE Xplore. Restrictions apply.
83
 2021 30th Wireless and Optical Communications Conference (WOCC)

(a) (b) (c)

Fig. 3. Performance versus number of devices. (a) Social Welfare. (b) Serving Ratio. (c) Utility of MEC layer. (d) Utility of Cloud layer. (e) Total Processing
Time(ms)

(2) Proposed- Edge: Edge-only system with enable pipeline- accuracy and latency with our management method and apply
sharing. it in several applications.
(3) Dedicated-C: Completely framework with a dedicated
V. C ONCLUSIONS
pipeline.
(4) Dedicated-Edge: Edge-only system with a dedicated We provided multi-application services on the system, such
pipeline. as facial recognition, object detection and pose estimation,
The proposed approaches enable one pipeline to serve and also proposed the RAMME method to deal with more
several task types. The dedicated serve by dedicated task scenarios and deploy with the intelligent MEC orchestrator.
type. We implemented the system with the CHT campus 5G system
(5) Greedy for Latency: Greedy for the minimum latency and run the complete service on our actual testbed. Finally,
configuration. we proved that the proposed scheme of the RAMME has a
(6) Greedy for Accuracy: Greedy for the highest accuracy better performance than other schemes and with higher social
configuration. welfare as well.

B. Implementation Results R EFERENCES

Figure 5 shows the performance with a different number of [1] S. Kekki, W. Featherstone, Y. Fang, P. Kuure, A. L. A. Ranjan,
D. Purkayastha, F. Jiangping, D. Frydman, G. Verin, K.-W. Wen, K. Kim,
devices in one request. We make each request with 1 to 10 R. Arora, A. Odgers, L. M. Contreras, and S. Scarpina, “Mec in 5g
devices simultaneously, and we will test the same number of networks,” ETSI White Paper, no. 28, 2019.
device requests 5 times with different user simulation parame- [2] Y. Chiang, C.-H. Hsu, and H.-Y. Wei, “Collaborative social-aware and
qoe-driven video caching and adaptation in edge network,” IEEE Trans-
ters. They will be performed on the above six schemes. Figure actions on Multimedia, pp. 1–1, 2020.
5(a) shows the proposed scheme has better social interest than [3] C. W. Y. Hung and R. Hwang, “Combinatorial clock auction for live
the other schemes. And compare to the two greedy schemes, video streaming in mobile edge computing,” IEEE INFOCOM 2018 -
IEEE Conference on Computer Communications Workshops (INFOCOM
the greedy of accuracy is higher than the greedy of latency WKSHPS), pp. 196–201, 2018.
because, in our calculation, the user quality may impact the [4] M. Aledhari, R. Razzak, R. Parizi, and F. Saeed, “Federated learning:
revenue and the welfare. However, the greedy of accuracy A survey on enabling technologies, protocols, and applications,” IEEE
Access, vol. 8, pp. 1–1, 01 2020.
method is still worse than our completely Proposed scheme. [5] J. Chen, K. Li, Q. Deng, K. Li, and P. S. Yu, “Distributed deep learning
Next, we find out these two greedy schemes can serve most model for intelligent video surveillance systems with edge computing,”
devices, except the requirements exceed the system’s limit IEEE Transactions on Industrial Informatics, pp. 1–1, 2019.
[6] Y. Huang, F. Wang, F. Wang, and J. Liu, “Deepar: A hybrid device-edge-
in Figure 5(b). Even if they can make most requests, they cloud execution framework for mobile deep learning applications,” IEEE
still perform better than our proposed scheme in Figure 5(a). Conf. Comput. Commun. Workshops (INFOCOM WKSHPS), p. 892–897,
That’s because it only considers quality or latency. Yet, there 2019.
[7] M. Liu, F. R. Yu, Y. Teng, V. C. M. Leung, and M. Song,
is a trade-off problem between accuracy and latency. Figure “Distributed resource allocation in blockchain-based video streaming
5(c) presents greedy for accuracy scheme costs more latency systems with mobile edge computing,” Trans. Wireless. Comm.,
than the other schemes. It only persuades the better quality vol. 18, no. 1, p. 695–708, Jan. 2019. [Online]. Available:
https://doi.org/10.1109/TWC.2018.2885266
but does not care about the processing time anymore. In [8] T.-Y. Kan, Y. Chiang, and H.-Y. Wei, “Task offloading and resource
contrast, greedy for latency only manage the processing time; allocation in mobile-edge computing system,” in 2018 27th Wireless and
as a result, it may perform in lower latency than others. The Optical Communication Conference (WOCC), 2018, pp. 1–4.
[9] Y. Zhang, J.-H. Liu, C.-Y. Wang, and H.-Y. Wei, “Decomposable in-
processing time with all of the schemes will increase with the telligence on cloud-edge iot framework for live video analytics,” IEEE
growth in the number of the device. Internet of Things Journal, vol. 7, no. 9, pp. 8860–8873, 2020.
In conclusion, our Proposed-C scheme performs the better
result in the testing demo. It may balance the trade-off between

Authorized licensed use limited to: VTU Consortium. Downloaded on October 22,2024 at 05:07:11 UTC from IEEE Xplore. Restrictions apply.
84