Content uploaded by Rashid Muhammad Sulaiman
Author content
All content in this area was uploaded by Rashid Muhammad Sulaiman on Mar 14, 2024
Content may be subject to copyright.
Graph neural network for in-network placement of real-time
metaverse tasks in next-generation network
Sulaiman Muhammad Rashida, Ibrahim Aliyua, Il-Kwon Jeongb, Tai-Won Umc,∗, Jinsul Kima,∗
aDepartment of Intelligent Electronics and Computer Engineering, Chonnam National
University, Gwangju, 61186, Korea
bHyper-Reality Metaverse Research Laboratory Content Research Division, Electronics and Telecommunications
Research Institute(ETRI), Daejeon, Korea
cSchool of Data Science, Chonnam National University, Gwangju, 61186, Korea
Abstract
This study addresses the challenge of real-time metaverse applications by proposing an in-
network placement and task-offloading solution for delay-constrained computing tasks in next-
generation networks. The metaverse, envisioned as a parallel virtual world, requires seamless
real-time experiences across diverse applications. The study introduces a software-defined net-
working (SDN)-based architecture and employs graph neural network (GNN) techniques for
intelligent and adaptive task allocation in in-network computing (INC). Considering time con-
straints and computing capabilities, the proposed model optimally decides whether to offload
rendering tasks to INC nodes or edge server. Extensive experiments demonstrate the superior
performance of the proposed GNN model, achieving 97% accuracy compared to 72% for mul-
tilayer perceptron (MLP) and 70% for decision trees (DTs). The study fills the research gap in
in-network placement for real-time metaverse applications, offering insights into efficient ren-
dering task handling.
Keywords: Delay-constrained computing, Graph neural network, In-Network computing,
Metaverse, Real-time rendering
1. Introduction
The Metaverse represents a hypothetical “parallel virtual world,” embodying lifestyles and
work environments within virtual cities as an alternative to future smart cities [1]. It envisions
seamless integration into our daily lives, offering users immersive real-time experiences globally
[2, 3, 4]. However, realizing this vision presents a substantial challenge in ensuring the metaverse
operates seamlessly in real-time, irrespective of geographical distances [5]. To achieve high
levels of immersion and realism, metaverse applications must render graphics, animations, and
simulations in real time [6] which is crucial for virtual reality (VR) and augmented reality (AR)
applications aiming to induce a sense of presence and “being there” [7]. In the metaverse, users
can generate and modify 3D digital content, where rendering plays a pivotal role in producing
∗Corresponding author
Email addresses: stwum@jnu.ac.kr (Tai-Won Um), jsworld@jnu.ac.kr (Jinsul Kim)
Preprint submitted to Internet of Things March 4, 2024
arXiv:submit/5446724 [cs.DC] 4 Mar 2024
such content [8], involving the computation of images perceived by users based on computer
graphics principles [9]. Therefore, efficient handling of rendering tasks for real-time metaverse
applications is crucial and imperative for research and development.
While mobile edge computing (MEC) offers a remedy through remote task offloading (TO), it
struggles to meet extensive user concurrent demands [10, 11, 12, 13]. The computing in the net-
work (COIN) paradigm emerges as a promising solution, utilizing untapped network resources
to execute tasks, diminishing latency, and fulfilling quality of experience (QoE) requirements
[10, 14]. Nonetheless, augmenting computing resources or enabling COIN leads to heightened
power consumption. Effectively allocating COIN computing resources in real-time to adapt to
continually shifting user demands while ensuring overall system availability poses a critical chal-
lenge.
1.1. Motivation and Contributions
Despite substantial advancements in handling time-intensive tasks, challenges persist, includ-
ing potential bottlenecks arising from resource constraints and network congestion [15]. Previ-
ous researchers propose the concept of in-network computing (INC) as a solution for managing
delay-constrained tasks [16, 17, 18]. This leverages the computing capability of nodes deployed
across the INC paradigm. Deciding which INC node should execute a task or whether to of-
fload it to the edge server due to network load can be achieved by adopting a software-defined
networking (SDN) architecture [19]. SDN provides programmability, flexibility, and centralized
control, enabling efficient management of network resources [20]. In this context, our primary
effort focuses on ensuring users experience seamless and immersive interactions within the meta-
verse while ensuring rendering tasks meet their deadlines. Rendering high-quality 3D content is
computationally intensive, requiring substantial processing power to handle the complex calcu-
lations [21]. It involves input processing, environment rendering, spatial mapping and tracking,
and composition and display [22, 9]. In this study, we consider the first three steps as one inde-
pendent task that can be computed by one node while another node executes the last step.
However, achieving optimal decisions requires complex mathematical models, posing a chal-
lenge to efficient task allocation. As outlined in [23], machine learning (ML) approaches can be
devised to determine how to allocate computing tasks across the network. This methodology
holds promise in dynamically adapting to the complexities of task allocation, offering a data-
driven framework to enhance real-time decision-making.
Recently, graph neural networks (GNN) have received increasing attention and are widely
utilized in various applications [24, 25, 26]. Their ability to model and analyze complex relation-
ships within graph-structured data makes them particularly promising for addressing rendering
task placement decisions. Our study will leverage graph convolutional networks (GCNs) and
develop a model capable of learning the dynamicity of the network and computing requests to
facilitate intelligent and adaptive rendering task allocation strategies.
Early studies mainly focused on optimizing rendering algorithms to address the challenges
of real-time rendering in metaverse applications, while other techniques, such as INC, which are
drawing more researchers’ attention in delay-constrained computing, are yet to be devised for
such problems. Previous works on INC focused on offloading tasks independently on SDN-based
edge networks; none considered computing tasks on more than one node due to the complexity of
being computed by a single node. Finally, different ML techniques have been utilized by various
researchers to address dynamic task placement problems in SDN edge environments. However,
GNN algorithms have not been applied to this field yet. Therefore, this study establishes a novel
2
approach for studying placement and offloading decisions for delay-constrained tasks for real-
time metaverse applications. This approach promises to employ and compare GNN with other
ML techniques to efficiently distribute rendering tasks within an INC environment and impact
the real-time rendering performance of metaverse applications.
The extensive literature reviewed in this study indicates that no works on in-network place-
ment of real-time rendering tasks in metaverse applications have been conducted. Specifically,
we consider a metaverse concert scenario where concert attendees gather to collectively interact
and experience the concert. The audience is dispersed worldwide, yet everyone is immersed in
the same live performance. In this context, the efficient distribution of real-time rendering tasks
within the metaverse network is crucial for providing concert attendees with a seamless and im-
mersive experience. The challenges of coordinating rendering tasks in a distributed and global
environment necessitate a specialized system tailored to the unique requirements of metaverse
concerts. This study introduces a groundbreaking approach to address this gap, presenting a
novel system designed to enhance the in-network placement of real-time rendering tasks. More
specifically, our system provides the following:
•We propose an SDN-based network architecture that includes the control and data planes
(comprising INC nodes, edge servers, and ingressors). The control plane receives requests
from ingressors and orchestrates communication with the selected INC node/edge server
for task execution.
•We formulate the optimal task placement problem through an Integer Linear Programming
(ILP) to minimize rendering latency under task time execution constraints, queuing delays,
and the limited computing capabilities of the executing nodes.
•We designed and implemented a GCN model to identify the near-optimal placement of
metaverse rendering tasks in real-time computing nodes, minimizing resource use while
meeting delay requirements.
•We conduct extensive experiments to evaluate the performance of the model results against
some benchmark solutions. The performance of our proposed model surpasses other
schemes in the context of delay-constrained computing in metaverse applications.
The rest of the paper is organized as follows: Section 2 addresses related studies and Section
3 presents the system model. In Section 4, we present optimal task placement using GNN;
Section 5 explains how we obtain our dataset; and Section 6 presents the numerical investigation
to validate the results of our proposed study. Conclusion and research directions are discussed in
Section 7.
2. Related work
The problem of delay-constrained computing has garnered considerable attention in recent
years [18, 27, 28]. Related research can be categorized into three main areas: first, there is a
focus on the in-network placement of delay-constrained tasks. Subsequently, attention is given
to the influence of ML in SDN-based task placement. Finally, contemporary studies concerning
real-time rendering in the metaverse are presented.
3
2.1. In-network placement of delay-constraint tasks
The future of computing is poised at the edge [29, 30], where data-driven decisions necessi-
tate near-instantaneous processing. In this scenario, INC nodes serve as execution units, dynam-
ically selected based on each task’s specific needs and deadlines. Extensive research has delved
into the task placement problem within INC environments [18, 31, 32, 33, 34]. Some approaches
involve offloading tasks to the INC node close to the user [34]. However, the resulting heavy
workload often increases waiting times, undermining the core objective of near-instantaneous
processing. Researchers are actively exploring potential solutions to mitigate this challenge.
For instance, notable work in [18], addresses the optimal task placement problem for delay-
constrained tasks by formulating it as a mixed-integer linear programming challenge. This line of
inquiry has been extended in [23] where supervised ML techniques are leveraged to enhance task
placement decisions. As these research efforts unfold, they pave the way for a more responsive
computing paradigm that aligns with the demands of data-driven, edge-centric decision-making.
Optimizing task placement in INC at the edge, especially for delay-constrained tasks, is crucial
given that heavy workloads can lead to increased waiting times, thereby impacting the goal of
achieving near-instantaneous processing for data-driven, edge-centric decision-making.
2.2. Impact of ML in SDN-based task placement
The separation of the control and data planes facilitated by SDN has introduced a dynamic
technique that enhances network flexibility [35]. Among researchers, ML approaches for task
placement in SDN computing are gaining traction [36, 37, 38]. For instance, in [37] authors
proposed a deep learning-based resource allocation within SDN-enabled Fog architecture, where
ML techniques are employed to compute resource links alongside other controller-imposed con-
straints. As SDN evolves, the integration of ML techniques optimizes task placement and en-
hances the adaptability and responsiveness of networks to dynamic conditions. The work pre-
sented in [23] signifies a pivotal shift toward intelligent and data-driven decision-making within
network environments. Various supervised learning algorithms, such as decision trees (DTs),
bagged trees, multilayer perception (MLP), and support vector machines (SVM), were utilized
to facilitate less complex and faster placement decision-making within an SDN infrastructure.
Given the frequent fluctuation in network conditions and requirements, AI models must strive to
deliver near-optimal task placement decisions.
2.3. Contemporary works in real-time rendering in the metaverse
As the metaverse gains prominence as a shared digital space where users engage with each
other and digital entities in real time, the demand for real-time rendering with stringent delay
constraints becomes imperative. Researchers are exploring innovative solutions to tackle the
challenges posed by rendering complex virtual environments in the metaverse while adhering
to strict time constraints [39, 40].Traditionally, rendering algorithms such as view frustum and
occlusion culling have been utilized by researchers to selectively remove objects outside the
camera’s view, thereby enhancing the performance of metaverse applications [41, 42]. However,
the need for innovative approaches becomes paramount as the metaverse evolves, with users
freely placing diverse objects. Recent works, as highlighted in [42], delve into object-culling
techniques that leverage vertex chunks to render a substantial volume of objects in real time. This
method compresses bounding boxes into data units, reducing input data for rendering passes. The
metaverse is characterized by dynamic and interactive environments [43]. Therefore, it is crucial
to explore new directions and methods that can adapt to changes in the virtual world in real time
and effectively handle dynamic scenarios without compromising performance.
4
Edge Server 1 Controller
Concert
Area
End User
Devices
INC Programmable
nodes
Task Producers
Players
Edge Server 2
Figure 1: Experimental Setup
3. System Model
This section offers an in-depth exploration of the framework that facilitates the utilization of
SDN-based in-network architecture to facilitate the placement and offloading of delay-constrained
tasks within real-time metaverse applications, focusing on the metaverse concert use case.
3.1. Network Model
Offloading all tasks to a single-edge server or node is suboptimal, as it can result in substantial
latency, which is unacceptable for delay-constrained applications and may lead to frequent dead-
line misses [44, 45, 46].As illustrated in Figure 1, we consider a metropolitan network system
comprising the control plane, data plane, ingressors, and concert area to mitigate this issue.
In our scenario, ingressors(group of players) i=1,2, ....Irequest a set of rendering tasks.
The Control Plane C P is crucial in orchestrating and managing overall network operations. It
acts as the system’s brain, making decisions based on real-time data and communication with
the network elements[47, 48]. The CP initializes resource information for each INC node i∈E,
which includes computing capability Ci, current network load NL, and features associated with
each rendering task r∈R. Upon receiving a rendering request, the CP returns the addressArof
selected nodesNrand establishes a communication link with nodes responsible for task execu-
tion. Tasks are divided between two INC nodes, denoted as node1 and node2, or else sent to the
edge server.
3.2. Task Model
All INC nodes, denoted by i∈E, possess computing capability Cimeasured in CPU cycles
per second. The ingressors request a set of rendering tasks R, characterized by (Sr,Pr,Mdel ),
where Srrepresents the size of task rin megabytes, Prdenotes the processing power required
5
Table 1: Symbol Descriptions
Symbol Description
Rset of rendering tasks
ESet of INC nodes
CiComputing capacity of node i
SrSize of rendering task r(mb)
PrRequired processing power for rendering task r
Mdel Maximum delay constraint for rendering task r
TrTotal time to execute task r
Cdel Total computation delay for task r
¯
CAverage computation time
hmin
1Minimum distance between ingressors and node1
hmin
2Minimum distance between node1 and node 2
αrArrival rate of task r
in CPU cycles to accomplish the task, and Mdel signifies the maximum delay constraint of the
rendering task.
The total time Tirequired by computing nodes to complete the task is given as
Ti=Pr
Ci1+Ci2
,i∈E(1)
where Ci1and Ci2represent the computing capabilities of the first and second INC nodes, re-
spectively.
Consistent with previous studies [19, 23], the arrival rate of requests for rendering task rat
node ifollows a Poisson distribution characterized by αr. The allocation of computing resources
received per task has an average distribution denoted by ¯
C; Consequently, all INC nodes can col-
lectively form an M/M/1 queuing model to process rendering tasks [49]. The total computation
delay for task rat nodes i1 and i2 will be:
Cdel =1
Ci1+Ci2
¯
C−Pi∈Rαr,i1,i2∈E(2)
3.3. Computing Model
The total network cost NC for a set of rendering task is determined by the distance of nodes
between the ingressors and the node1 as hmin
1, the distance of nodes between node1 and node2 as
hmin
2, and the size Siof the task, it can be derived as:
NC =αr
R
X
r=1
Si(hmin
1+hmin
2),i∈E;r∈R(3)
The number of exchanged nodes substantially impact the network cost, potentially resulting
in high traffic within the network. Let X1iand X2idenotes binary decisions where ’1’ signifies
that the rendering task is executed by node 1 and 2, respectively; otherwise, ’0’. Consequently,
we formulate the problem as follows:
6
AR/VR app Control Plane Data Plane
Rendering Request
Statistics on Resources
IP Address of Selected
Nodes
Request to run In-network Computing Program
Figure 2: Sequence diagram of rendering request and INC procedures
P: min X
i∈E
X
r∈R
X1iX2iαr
r
X
r=1
Si(hmin
1+hmin
2) (4)
Constraints:
X
i∈E
X1i+X
i∈E
X2i≤2 (5)
X
i∈E
X1i≤1 (6)
X1i+X2i≥1∀i∈E(7)
X1X2¯
C≤X1X2Mdel (8)
Ci
¯
C−X
r∈R
αr>0 (9)
X1,X2∈ {0,1},i∈E,r∈R
Constraints (5) and (6) described by nonlinear inequalities, ensure that a maximum of two nodes
executes each rendering task and only one node executes at a time. Equation (7) ensures that at
least one node is chosen for the task. Equations (8) and (9) confirm that the computing resources
allocated to a rendering task are within the limits. Table 2 presents the simulation parameter
settings except otherwise stated.
7
4. Optimal task placement using GNN
The ILP problem in eq. (4) corresponds to the Generalized Assignment Problem (GAP) [50,
23], which is NP-hard problem. Due to the complexity of rendering task placement decisions
under the constraint capacity of INC nodes and computation delay deadlines, the offloading and
placement problem is challenging, as it is difficult to predict task duration. Therefore, a more
dynamic solution is needed to adapt to delay and queuing time changes.
The decision-making process in real-time metaverse applications, especially regarding ren-
dering tasks, involves numerous variables and dependencies. Traditional approaches fall short
of capturing the complexity of these interactions and adapting to the dynamic nature of network
conditions [1, 51].Neural networks offer a compelling solution for several reasons: Non-linearity
handling [52], graph representation [53], and adaptability to dynamic environments.
Table 2: Parameter settings
Parameter Setting
Computing capability (Ci)
•Middle layer nodes: 5 ×108CPU cycles /s
•Upper layer nodes: 109CPU cycles /s
•Edge servers: 1010 CPU cycles /s
Size of rendering task (Sr) 10 MB
Average processing power required (Pr) 107
Maximum delay constraint (Mdel ) Uniformly distributed [10, 150] ms
Arrival rate of task (αr) 10 requests/s
We propose GNN approach, specifically GCN to train on labeled data derived from our op-
timal decisions based on historical scenarios. The inputs to the model include the total network
cost for each rendering task on all INC nodes, the maximum computation delay of the task, and
the Boolean decision that indicates whether the node satisfied the task constraint.
For each task run, each algorithm returns output with the selected nodes as the executors of
the rendering task. Once the executor is selected, the controller returns the address of the selected
nodes, and the task is sent to the executors for rendering task execution.
We first utilize a unified formula to define a graph, G=(V,E,X), where Vrepresents the set
of nnodes and Ethe set of edges. In the context of this study, Vsignifies the tasks and source
hosts, while Erepresents the connections/links between tasks and ingressors. The feature matrix
Xcontains the feature vectors associated with each node. We conducted several fine tuning and
hyperparameter tests to determine the best parameters for the model (see Table 4).
G=(V,E,X) (10)
V={vi,...,vn}(11)
E={(vi,vj)}(12)
Here, viand vjrepresent the k-th element of the task and source vectors of the set of rendering
requests R, respectively.
8
Input
Features Output
GV(X,W1)GV(H,W2)
Candidate
executor
Candidate
executor
address
ReLu
Network
Cost
Maximum
Computation
Delay
Boolean
Decision
Training
and
Inference
Model
Function
Figure 3: Graph convolutional model
X=[x1,x2,...,xn],where xirepresents the feature vector of node i.(13)
The model comprises two graph convolution layers: GV(X,W1) and GV(H,W2). Here, His
the hidden layer output, and W1and W2denote the weight matrices associated with the graph
convolution layers.
H0=X(14)
Hi=ReLU(GV(Hi−1,Wi)),i=1,...,n−1 (15)
Y=GV(Hn−1,Wn) (16)
The parameters of the GNN model are denoted by θ. The objective is to minimize the cross-
entropy loss function.
min θL(GV(. . . (ReLU(GraphConv(X, θ1)), θ2)...,θn−1), θn) (17)
The model computes logits ˆ
Y, representing the predicted labels, where ˆ
Y=model(G,X).
The loss is then computed as L(ˆ
Y,Ytrain), where Ytrain is the ground truth for the training set.
5. Dataset generation
Our experiment utilized a dataset generated by the optimal solution from Section 3 through
a standard optimization solver (Gurobi). The dataset comprises 100,000 samples (task requests).
To create training and validation sets, we employed the 10-fold cross-validation method, dividing
the dataset into 10 folds using random sampling without repetition. This method involved 10
training sessions, each utilizing a different fold as a validation set and the remaining folds as
training sets. The prediction error was determined as the mean average of the 10 individual
errors from each training session, ensuring an unbiased performance evaluation of the model
across different dataset partitions.
9
6. Numerical Investigation
In this section, we present the numerical result of the proposed model and compare its per-
formance with that of other models from the literature.
Table 3: Hyperparameter settings for training of GNN
Setting Value
Model type Graph Convolutional Network (GCN)
Number of features 30
Activation function ReLU
Layers 2
Optimizer Adam
Epochs 200
Learning rate 0.01
6.1. Experiment Environment
We utilized an SDN network topology comprising 30 INC nodes, as illustrated in Figure 1.
Four nodes transmit tasks to the edge servers. In contrast, upper-layer nodes are interconnected
in a fully meshed topology, serving as the root of a binary tree spanning three layers. The lowest
layer nodes, i.e., the leaf nodes of the tree, function as the ingressors.
The following settings obtain the Optimal solution: The simulation and model training are
being run using Spyder (Integrated Development Environment) IDE for Python programming
language, and the solver used is Gurobi 11.0.0 and Pyomo framework.
Table 4: GNN Hyperparameter selection results
Epochs Optimizer Activation function Layers Accuracy F1-score
50 Adam ReLU 2 0.82 0.81
100 Adam ReLU 2 0.93 0.92
200 Adam ReLU 2 0.97 0.96
50 Adam Sigmoid 2 0.68 0.58
100 Adam Sigmoid 2 0.71 0.60
200 Adam Sigmoid 2 0.70 0.59
50/300 SGD ReLU 2 0.21 0.04
50/300 SGD Sigmoid 2 0.65 0.53
50 RMSprop ReLU 2 0.43 0.31
200 RMSprop ReLU 2 0.61 0.50
100 Adamax ReLU 2 0.89 0.86
200 Adamax ReLU 2 0.94 0.92
100 Adam ReLU 3 (64, 128, 256) 0.93 0.91
200 Adam ReLU 3 (64, 128, 256) 0.86 0.82
10
6.2. Algorithm Computation time analysis
As depicted in Figure 4, the time complexity of algorithm computation for the ML models
compared to the optimal solution demonstrates that the optimal solution computes approximately
100 times faster than the ML models. Moreover, this difference substantially increases as the
number of rendering requests escalates. The simulation was executed on a 12th Gen Intel(R)
Core(TM) i5-12400F, 2.50 GHz (CPU), 16 GB (RAM), and 500GB HD.
Figure 4: Time complexity analysis of Algorithm computation of the ML models as compared to the optimal solution
6.3. GNN vs other ML models
It can be observed in Figure 5a, that the percentage of data offloaded to the edge server in the
investigated GNN model is relatively similar to that of the optimized solution. This observation
indicates that tasks whose deadlines cannot be met will be forwarded to the cloud, preventing
network congestion and reducing the risk of missing task deadlines. However, compared to
previous studies’ approaches (DT and MLP), they exhibit a slightly higher percentage of tasks
being offloaded to the edge server than GNN, resulting in more data being exchanged within the
network.
GNN demonstrates efficiency in managing resource costs when running sets of rendering
tasks. As shown in Figure 5b, there is only a nearly 2% increase in the resources used across all
requests compared to MLP’s 18% and DT’s 19% against the optimal solution. This finding also
indicates that a set of rendering requests will cover slightly less distance in GNN, resulting in
lower latency than the other models.
6.4. Performance relative to Baseline Scheme
The baseline scheme is the cloud-based approach, where all rendering tasks are offloaded to
the remote cloud for execution. As the number of rendering tasks increases, the total amount of
exchanged data within the domain increases, as shown in Figure 6a. The cloud-based approach
11
(a) Task offload to edge server
100 200 300 400 500
0
400
800
1200
1600
2000
Network cost (No of resources)
Number of requests
Optimal Solution
GNN
MLP
DT
(b) Network cost in terms of resources
Figure 5: GNN vs the Optimal Solution and other ML models
exhibits substantially more input data passing within the network from ingressors to the candidate
executor, resulting in higher resource usage and data exchange. In contrast, GNN demonstrates
less data exchange with fewer rendering requests, outperforming other models as the network
becomes overwhelmed with rendering requests.
The percentage of tasks being offloaded to the cloud affects the execution time of the tasks,
as depicted in Figure 6b. Since the cloud is expected to be a powerful executor with unlimited
computing capability, the cloud-based approach exhibits very low execution time compared to
other solutions. Nevertheless, GNN, compared to other ML models, requires less time for task
execution.
(a) Exchanged data (b) Average execution
Figure 6: Performance relative to Baseline Scheme
6.5. GNN vs other ML models in terms of delay constrained
We compare the performance of GNN relative to MLP and DT, focusing on the impact of
delay constraints. Figure 7a shows that the time required to execute tasks with delay constraints
12
is lower in every aspect. This is because even if the executor’s capability meets the task’s re-
quirements, the controller also checks the time constraint. If the time constraint cannot be met,
the task is forwarded to the edge server to avoid congestion within the network.
(a) Average execution time (b) Average queuing time
(c) Average computation time
Figure 7: Impact of delay constrained
Figure 7b illustrates the average time it takes for a set of tasks to wait before execution. Tasks
take longer waiting in the queue when no delay constraint is attached to the task. In Figure 7c, we
compute the average computation time, which sums the execution and queuing time, representing
the overall time it takes to compute set of rendering task.
6.6. ML related performance
We compared our proposed model with previous supervised ML models used in the literature
[17], and the metrics demonstrate that GNN outperforms all the models. . Regarding accuracy,
the ratio of correctly predicted observations to the total number of observations, GNN correctly
predicted 97% of all instances, surpassing MLP’s 72% and DT’s 70%. The output from the
optimized solution results in class imbalance due to the infrequent rendering tasks sent to the
cloud. Unlike GNN, this imbalance affects the metrics (precision, recall, F1-score) of DT and
MLP, resulting in poor performance, as evident in Table 5.
13
Table 5: Comparison of Metrics for GNN, MLP, and DT
Metrics GNN MLP DT
Accuracy 0.97 0.72 0.70
Precision 0.96 0.65 0.47
Recall 0.97 0.68 0.52
F1-score 0.96 0.66 0.48
7. Conclusion and Future work
Considering the heavy training demands, there will be limited capabilities for the INC nodes
during task computation. Since the task allocation problem is an online problem, it cannot be
efficiently solved by standard optimization solvers. Therefore, we employed GNN to solve this
work’s formulated task placement problem offline.
In this study, we proposed a GNN model for efficiently placing in-network tasks in the meta-
verse, addressing rendering tasks’ complexity and time sensitivity against edge limitations. Com-
pared to the baselines, MLP, and DT, the GNN substantially improved accuracy and performance
for real-time rendering tasks. Our model substantially impacts real-time metaverse applications
by integrating INC with edge computing and SDN architecture for efficient task allocation. Fu-
ture studies should enhance the model’s scalability and robustness, vital for its effectiveness in
extensive, complex metaverse settings.
Acknowledgment
This research was suported in part by the Culture, Sports and Tourism R&D Program throgh
the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism
in 2022 (Project Name: Development of real-time interactive metaverse performance experi-
ence platform technology on the scale of a large concert hall Project Number: RS-2022-050002,
Contribution Rate: 50%) and in part by Innovative Human Resource Development for Local
Intellectualization program through the Institute of Information & Communications Technol-
ogy Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (IITP-2024-
RS-2022-00156287, 50%)
References
[1] Z. Allam, A. Sharifi, S. E. Bibri, D. S. Jones, J. Krogstie, The metaverse as a virtual form of smart cities: Oppor-
tunities and challenges for environmental, economic, and social sustainability in urban futures, Smart Cities 5 (3)
(2022) 771–801. doi:https://doi.org/10.3390/smartcities5030040.
[2] R. Chengoden, N. Victor, T. Huynh-The, G. Yenduri, R. H. Jhaveri, M. Alazab, S. Bhattacharya, P. Hegde, P. K. R.
Maddikunta, T. R. Gadekallu, Metaverse for healthcare: A survey on potential applications, challenges and future
directions, IEEE Access 11 (2023) 12765–12795. doi:https://doi.org/10.1109/ACCESS.2023.3241628.
[3] Z. Ramadan, Marketing in the metaverse era: Toward an integrative channel approach, Virtual Reality 27 (2023)
1905–1918. doi:https://doi.org/10.1007/s10055-023- 00783-2.
[4] D. Buhalis, D. L. b, M. Lin, Metaverse as a disruptive technology revolutionising tourism management and mar-
keting, Tourism Management 97 (2023). doi:https://doi.org/10.1016/j.tourman.2023.104724.
[5] M. Xu, W. C. Ng, W. Y. B. Lim, J. Kang, Z. Xiong, D. Niyato, Q. Yang, X. Shen, C. Miao, A full dive into realizing
the edge-enabled metaverse: Visions, enabling technologies, and challenges, IEEE Communications Surveys &
Tutorials 25 (1) (2023) 656–700. doi:https://doi.org/10.1109/COMST.2022.3221119.
14
[6] S. Cheng, Metaverse and immersive interaction technology, in: Metaverse: Concept, Content and Context, 2023.
doi:https://doi.org/10.1007/978-3- 031-24359-2.
[7] S. Dhelim, T. Kechadi, L. Chen, N. Aung, H. Ning, L. Atzori, Edge-enabled metaverse: The convergence of
metaverse and mobile edge computing, IEEE IoT Journal Distributed, Parallel, and Cluster Computing (cs.DC)
14 (8) (2022). doi:https://doi.org/10.48550/arXiv.2205.02764.
[8] D. Liu, L. Wei, Q. Zheng, P. Ding, Y. Shen, Design and implementation of dis-
tributed rendering system, in: Haikou, China, 2022. doi:https://doi.org/10.1109/
SmartWorld-UIC- ATC-ScalCom-DigitalTwin-PriComp-Metaverse56740.2022.00332.
[9] A. Tewari, J. Thies, B. Mildenhall, P. Srinivasan, E. Tretschk, W. Yifan, C. Lassner, V. Sitzmann, R. Martin-Brualla,
S. Lombardi, T. Simon, C. Theobalt, M. Nießner, J. T. Barron, G. Wetzstein, M. Zollh ¨
ofer, V. Golyanik, Advances in
neural rendering, Computer Graphics Forum 41 (2) (2022) 703–735. arXiv:https://onlinelibrary.wiley.
com/doi/pdf/10.1111/cgf.14507,doi:https://doi.org/10.1111/cgf.14507.
URL https://onlinelibrary.wiley.com/doi/abs/10.1111/cgf.14507
[10] I. Aliyu, S. Oh, N. Ko, T.-W. Um, J. Kim, Dynamic partial computation offloading for the metaverse in in-network
computing, IEEE Access 12 (2024) 11615–11630. doi:10.1109/ACCESS.2023.3344817.
[11] H. Jiang, X. Dai, Z. Xiao, A. Iyengar, Joint task offloading and resource allocation for energy-constrained mobile
edge computing, IEEE Transactions on Mobile Computing 22 (7) (2023) 4000–4015. doi:10.1109/TMC.2022.
3150432.
[12] A. Islam, A. Debnath, M. Ghose, S. Chakraborty, A survey on task offloading in multi-access edge computing, Jour-
nal of Systems Architecture 118 (2021) 102225. doi:https://doi.org/10.1016/j.sysarc.2021.102225.
URL https://www.sciencedirect.com/science/article/pii/S1383762121001570
[13] J. Chen, Y. Leng, J. Huang, An intelligent approach of task offloading for dependent services in mo-
bile edge computing, Journal of Cloud Computing 12 (107) (2023). doi:https://doi.org/10.1186/
s13677-023- 00477-9.
[14] H. Wu, Z. Xiang, G. T. Nguyen, Y. Shen, F. H. Fitzek, Computing meets network: Coin-aware offloading for
data-intensive blind source separation, IEEE Network 35 (5) (2021) 21–27. doi:10.1109/MNET.011.2100060.
[15] A. Shakarami, A. Shahidinejad, M. Ghobaei-Arani, A review on the computation offloading approaches in mobile
edge computing: A game-theoretic perspective, Software: Practice and Experience 50 (9) (2020) 1719–1759.
doi:https://doi.org/10.1002/spe.2839.
[16] S. Kianpisheh, T. Taleb, A survey on in-network computing: Programmable data plane and technology specific
applications, IEEE Communications Surveys & Tutorials 25 (1) (2023) 701–761. doi:https://doi.org/10.
1109/COMST.2022.3213237.
[17] T. Mai, H. Yao, S. Guo, Y. Liu, In-network computing powered mobile edge: Toward high performance industrial
iot, IEEE Network 35 (1) (2021) 289–295. doi:https://doi.org/10.1109/MNET.021.2000318.
[18] G. Lia, M. Amadeo, C. Campolo, G. Ruggeri, A. Molinaro, Optimal placement of delay-constrained in-network
computing tasks at the edge with minimum data exchange, in: 2021 IEEE 4th 5G World Forum (5GWF), Montreal,
QC, Canada, 2021. doi:https://doi.org/10.1109/5GWF52925.2021.00091.
[19] M. Priyadarsini, P. Bera, Software defined networking architecture, traffic management, security, and placement:
A survey, Computer Networks 192 (2021). doi:https://doi.org/10.1016/j.comnet.2021.108047.
[20] J. Xie, D. Guo, Z. Hu, T. Qu, P. L. b, Control plane of software defined networks: A survey, Computer Communi-
cations 67 (2015) 1–10. doi:https://doi.org/10.1016/j.comcom.2015.06.004.
[21] C.-H. Lin, J. Gao, L. Tang, T. Takikawa, X. Zeng, X. Huang, K. Kreis, S. Fidler, M.-Y. Liu, T.-Y. Lin, Magic3d:
High-resolution text-to-3d content creation, in: Computer Vision and Pattern Recognition, 2023.
[22] M. Huzaifa, R. Desai, S. Grayson, X. Jiang, Y. Jing, J. Lee, F. Lu, Y. Pang, J. Ravichandran, F. Sinclair, B. Tian,
H. Yuan, J. Zhang, S. V. Adve, Illixr: Enabling end-to-end extended reality research, in: IEEE International Sym-
posium on Workload Characterization (IISWC), Storrs, CT, USA, 2021. doi:https://doi.org/10.1109/
IISWC53511.2021.00014.
[23] G. Lia, M. Amadeo, G. Ruggeri, C. Campolo, A. Molinaro, V. Loscr`
ı, In-network placement of delay-constrained
computing tasks in a softwarized intelligent edge, Computer Networks 219 (2022). doi:https://doi.org/10.
1016/j.comnet.2022.109432.
[24] J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, M. Sun, Graph neural networks: A review of
methods and applications, AI Open 1 (2020) 57–81. doi:https://doi.org/10.1016/j.aiopen.2021.01.
001.
[25] F. Liang, C. Qian, W. Yu, D. Griffith, N. Golmie, Wireless communications and mobile computingAdvances of
Intelligent Sensory Data Processing and Protection in IoT (2022). doi:https://doi.org/10.1155/2022/
9261537.
[26] A. Gupta, P. Matta, B. Pant, Graph neural network: Current state of art, challenges and applications, Materials
Today: Proceedings 46 (20) (2021) 10927–10932. doi:https://doi.org/10.1016/j.matpr.2021.01.950.
[27] A. Bozorgchenani, F. Mashhadi, D. Tarchi, S. A. Salinas Monroy, Multi-objective computation sharing in energy
15
and delay constrained mobile edge computing environments, IEEE Transactions on Mobile Computing 20 (10)
(2021) 2992–3005. doi:10.1109/TMC.2020.2994232.
[28] R. Fantacci, B. Picano, Performance analysis of a delay constrained data offloading scheme in an integrated cloud-
fog-edge computing system, IEEE Transactions on Vehicular Technology 69 (10) (2020) 12004–12014. doi:
10.1109/TVT.2020.3008926.
[29] S. Deng, H. Zhao, W. Fang, J. Yin, S. Dustdar, A. Y. Zomaya, Edge intelligence: The confluence of edge computing
and artificial intelligence, IEEE Internet of Things Journal 7 (8) (2020) 7457–7469. doi:10.1109/JIOT.2020.
2984887.
[30] K. Cao, Y. Liu, G. Meng, Q. Sun, An overview on edge computing research, IEEE Access 8 (2020) 85714–85728.
doi:10.1109/ACCESS.2020.2991734.
[31] A. Munir, T. He, R. Raghavendra, F. Le, A. X. Liu, Network scheduling and compute resource aware task placement
in datacenters, IEEE/ACM Transactions on Networking 28 (6) (2020) 2435–2448. doi:10.1109/TNET.2020.
3013548.
[32] J. Gedeon, M. Stein, L. Wang, M. Muehlhaeuser, On scalable in-network operator placement for edge computing,
in: 2018 27th International Conference on Computer Communication and Networks (ICCCN), 2018, pp. 1–9.
doi:10.1109/ICCCN.2018.8487419.
[33] N. Hu, Z. Tian, X. Du, M. Guizani, An energy-efficient in-network computing paradigm for 6g, IEEE Transactions
on Green Communications and Networking 5 (4) (2021) 1722–1733. doi:10.1109/TGCN.2021.3099804.
[34] Y. Jin, H. Lee, On-demand computation offloading architecture in fog networks, Electronics 8 (10) (2019). doi:
10.3390/electronics8101076.
URL https://www.mdpi.com/2079-9292/8/10/1076
[35] N. M. Kaliyamurthy, S. Taterh, S. Shanmugasundaram, A. Saxena, O. Cheikhrouhou, H. B. Elhadj, Software-
defined networking: An evolving network architecture—programmability and security perspective, Security and
Communication Networks 2021 (2021). doi:https://doi.org/10.1155/2021/9971705.
[36] B. Sarma, R. Kumar, T. Tuithung, Machine learning enabled network and task management in sdn based fog
architecture, Computers and Electrical Engineering 108 (2023) 108705. doi:https://doi.org/10.1016/j.
compeleceng.2023.108705.
URL https://www.sciencedirect.com/science/article/pii/S0045790623001295
[37] A. Lakhan, M. A. Mohammed, O. I. Obaid, C. Chakraborty, K. H. Abdulkareem, S. Kadry, Efficient deep-
reinforcement learning aware resource allocation in sdn-enabled fog paradigm, Automated Software Engineering
29 (20) (2022). doi:https://doi.org/10.1007/s10515-021-00318-6.
[38] R. Amin, E. Rojas, A. Aqdus, S. Ramzan, D. Casillas-Perez, J. M. Arco, A survey on machine learning techniques
for routing optimization in sdn, IEEE Access 9 (2021) 104582–104611. doi:10.1109/ACCESS.2021.3099092.
[39] A. Hazarika, M. Rahmati, Towards an evolved immersive experience: Exploring 5g- and beyond-enabled ultra-low-
latency communications for augmented and virtual reality, Sensors 23 (7) (2023). doi:10.3390/s23073682.
URL https://www.mdpi.com/1424-8220/23/7/3682
[40] M. T. Vega, C. Liaskos, S. Abadal, E. Papapetrou, A. Jain, B. Mouhouche, G. Kalem, S. Erg¨
ut, M. Mach, T. Sabol,
A. Cabellos-Aparicio, C. Grimm, F. D. Turck, J. Famaey, Immersive interconnected virtual and augmented reality:
A 5g and iot perspective, Journal of Network and Systems Management 28 (2020) 796–826. doi:https://doi.
org/10.1007/s10922-020- 09545-w.
URL https://www.sciencedirect.com/science/article/pii/S2468502X22000158
[41] J. Xue, X. Zhai, H. Qu, Efficient rendering of large-scale cad models on a gpu virtualization architecture with model
geometry metrics, in: 2019 IEEE International Conference on Service-Oriented System Engineering (SOSE), 2019,
pp. 251–2515. doi:10.1109/SOSE.2019.00043.
[42] E.-S. Lee, B.-S. Shin, Vertex chunk-based object culling method for real-time rendering in metaverse, Electronics
12 (12) (2023). doi:https://doi.org/10.3390/electronics12122601.
[43] Y. Zhao, J. Jiang, Y. Chen, R. Liu, Y. Yang, X. Xue, S. Chen, Metaverse: Perspectives from graphics, interactions
and visualization, Visual Informatics 6 (1) (2022) 56–67. doi:https://doi.org/10.1016/j.visinf.2022.
03.002.
URL https://www.sciencedirect.com/science/article/pii/S2468502X22000158
[44] A. Naouri, H. Wu, N. A. Nouri, S. Dhelim, H. Ning, A novel framework for mobile-edge computing by optimiz-
ing task offloading, IEEE Internet of Things Journal 8 (16) (2021) 13065–13076. doi:10.1109/JIOT.2021.
3064225.
[45] A. Avan, A. Azim, Q. H. Mahmoud, A state-of-the-art review of task scheduling for edge computing: A delay-
sensitive application perspective, Electronics 12 (12) (2023). doi:10.3390/electronics12122599.
URL https://www.mdpi.com/2079-9292/12/12/2599
[46] L. Lov´
en, E. Peltonen, L. Ruha, E. Harjula, S. Pirttikangas, A dark and stormy night: Reallocation storms in edge
computing, J Wireless Com Network 86 (2022). doi:https://doi.org/10.1186/s13638- 022-02170-y.
URL https://www.mdpi.com/2079-9292/12/12/2599
16
[47] M. Paliwal, D. Shrimankar, O. Tembhurne, Controllers in sdn: A review report, IEEE Access 6 (2018) 36256–
36270. doi:10.1109/ACCESS.2018.2846236.
[48] B. Isong, T. Kgogo, F. Lugayizi, B. Kankuzi, Trust establishment framework between sdn controller and applica-
tions, in: 2017 18th IEEE/ACIS International Conference on Software Engineering, Artificial Intelligence, Net-
working and Parallel/Distributed Computing (SNPD), 2017, pp. 101–107. doi:10.1109/SNPD.2017.8022707.
[49] K. V. Vijayashree, B. Janani, Transient analysis of an m/m/1 queueing system subject to differentiated vacations,
Quality Technology & Quantitative Management 15 (6) (2018) 730–748. doi:https://doi.org/10.1080/
16843703.2017.1335492.
[50] D. G. Cattrysse, L. N. Van Wassenhove, A survey of algorithms for the generalized assignment problem, European
Journal of Operational Research 60 (3) (1992) 260–272. doi:https://doi.org/10.1016/0377- 2217(92)
90077-M.
URL https://www.sciencedirect.com/science/article/pii/037722179290077M
[51] S.-M. Park, Y.-G. Kim, A metaverse: Taxonomy, components, applications, and open challenges, IEEE Access 10
(2022) 4209–4251. doi:10.1109/ACCESS.2021.3140175.
[52] S. Pal, The role of neural networks in predicting supply chain disruptions, International Journal of All Research
Education and Scientific Methods (IJARESM) 11 (11) (2023) 134–140.
[53] J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, M. Sun, Graph neural networks: A review of
methods and applications, AI Open 1 (2020) 57–81. doi:https://doi.org/10.1016/j.aiopen.2021.01.
001.
17
