Page 198 - Kaleidoscope Academic Conference Proceedings 2021

2021 ITU Kaleidoscope Academic Conference

4.2.1 Resource allocation approaches

Considering the functions mentioned above, there are two
types of approaches for decision-making available:

1) Policy-based: In this approach, a list of prices for slices
and resources are provided by the Mobile Network
Operator (MNO) and the final decision is made
considering the current state of the system and the
network’s policy.
2) Auction-based: Here, there is no fixed price for the
slices or resources. Instead, the MNO provides a list of
available slices and resources, and tenants can request
a service with an offer bid. Then, the MNO chooses

between the tenants and provide the winner tenant
Figure 2 – The relationship between artificial intelligence with the requested slice or resource [20].
and orchestration functionalities
4.3 Operation & Management
capacity efficiently and minimize the operation costs. The
main challenge here is to keep a balanced allocation where In today’s network architecture, humans play a vital role in
none of the following happens: the stage of management and operation and by automating
network functions, network costs decrease significantly. In
1. Under Provisioning: If less resources are allocated to a order to manage a network successfully, monitoring is an
slice than needed, a Service Level Agreement (SLA) inseparable function to implement in all steps of network
set with the tenant would be violated. management and is an integral part of network management
2. Over Provisioning: if more resources than required are functions. The following information is obtained by
allocated to a slice, extra resources remain unused, monitoring. System reports, traffic analysis, resource status
making it a waste of resource and money [17]. spectral clustering, K-mean clustering, support vector
machine and deep neural network are ML techniques that
To find the proper state of resource allocation, future requests can be used for this task. The main purposes of these
for each slice should be predicted and also required resources functions are categorizing system operation and forecasting
for each service should be determined, which is a complex resource utilization [16]. Network slicing management can
task for operators and we cannot address them individually be put into several tasks. These tasks are performance
[18]. While traditional solutions aren’t helpful for this management, fault management and security.
purpose, artificial intelligence is useful to solve both aspects
of resource management and orchestration. In order to do 4.3.1 Performance management
that, we can implement a Convolutional Neural Network
(CNN) in which respective costs of SLA violations (under The first step in controlling the performance of a network is
provisioning) and the costs wasting the resources (over admission control. Admission control is about determining
provisioning) is considered as a loss function to minimize whether a network can accept the upcoming slice requests
[19]. The policies which determine the allocation of and provide the accepted requests with its requirements or
resources between slices should adapt to the dynamic and not. Having limited resources, admission control becomes a
changing behavior of the network slices. Take a situation vital role in slice management. According to the 3GPP (3rd
where we have emergency requests and other services for Generation Partnership Project) standardization of network
example. In this situation, when we have numerous requests slicing, tenants (the communication service client) send
coming to the network, due to the sharing of our limited requests for specific services and then based on some fixed
resources between these two services, the higher priority is factors, the cost is calculated and paid. Having vigorous
given to the former services [16]. The main purpose of these resource sharing, the network would not be able to meet
two first steps is to classify service requirements, predict Key Performance Indicators (KPIs), resulting in a decrease
network trends and user behavior and configure network in the network revenue since the required services are not
parameters. The machine learning techniques used to do provided. On the other way around, rejecting most of the
these tasks are support vector machine, gradient boosting incoming requests, the network would lose many
decision tree, spectral clustering and reinforcement learning. opportunities to gain profit. Therefore, one profound action
Classifying a new service into one of the 3 categories (eMBB, to be taken in this phase is to establish a balance between key
mMTC, and URLLC) should be done based on given factors performance indicators and resource sharing [17]. Another
and for this goal, other supervised and unsupervised point worth mentioning is that meeting the right KPIs needs
techniques can be used too. Moreover, to determine proper complete slice isolation, making it more complicated to find
network parameters, reinforcement learning is usually the equilibrium point. In order to overcome this issue,
helpful. Conclusively, with the help of these techniques and admission control should have the information about the
algorithms, we can design and construct an efficient sliced dynamic behavior of slices along with
network which can adjust to new services and use cases [16].

– 136 –