Page 91 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

Planning and design Construction and deployment
Requirements analysis,
environment analysis, Static resource allocation
topology determination VNF placement, orchestration

Fault detection Operation, control and management Monitoring
Syslog analysis, Dynamic resource allocation, Workload,
behavior analysis, adjustment; performance,
fault location policy adaptation resource utilization

Security
Traffic analysis, DPI,
threat identification
infection isolation

Figure 2. Functions for network automation.

requirements in terms of network performance metrics such Thanks to the technologies of NFV and SDN, each type of
as network latency, packet loss rate, and bandwidth. These IoT service can be deployed in a distinct network slice
requirements are mapped into a virtual topology containing configured with the appropriate amount of resources leased
the required amount of computing and networking from the substrate networks of one or several infrastructure
resources. The VNO then sends a network resource request providers. For example, autonomous cars, social
containing the logical topology and required performance networking services and sensors/actuators of environment
metrics to the infrastructure provider to lease the required and weather monitoring services in Fig. 1 are connected to
amount of virtualized resources. three different network slices. These slices are
independently configured, deployed, and operated either by
Infrastructure provider performs mapping of the logical the same or different virtual network operators.
network topology onto the virtualized resources of
substrate networks. This mapping process is also referred to Since most steps involved in construction of a
as virtual network embedding. Infrastructure provider telecommunication network currently require manual
allocates the requested amount of virtualized resources and operations, it takes about two weeks in most cases to
provides the corresponding resource control and complete the construction and start serving customers [6].
management interfaces to VNO. Similar amount of time would be required to construct a
network slice if the operations are carried out manually.
VNO installs necessary network functions, protocols, Therefore, automation technologies for the construction,
customer registry and software platform on the leased deployment, operation, monitoring, and control are
virtualized resources. It then deploys the virtual network essential to shorten the time for on-demand construction of
for the application service and regularly monitors the a network slice and enable faster roll out of new IoT
performance to check if currently allocated resources are in services. Automation technologies would also help in
appropriate amounts to execute the necessary network dealing with the shortage of skilled manpower or avoiding
functions to process the given workload. If it determines human errors in the tedious manual operations. The
that resources need to be adjusted, it sends resource autonomous functions carry out resource abstraction,
adjustment requests containing the amounts of required allocation, arbitration, adjustment, and adaptation. These
resources to infrastructure provider. Alternatively, functions enable network slices to adapt to dynamic
infrastructure provider may regularly monitor the utilization network environments (i.e., changing workload, customer
of virtualized resources and adjusts the resource amount base, and link quality). AI and machine learning based
according to the service level agreement made with VNOs. methods play a key role for enabling the automation of
To monitor and adjust resources, infrastructure provider network control and management.
uses network and node virtualization tools such as
OpenFlow and OpenStack. Open Networking Lab 3. NETWORK AUTOMATION FUNCTIONS
(onlab.us) to develop open and innovative technology
called Open Network Operating System (ONOS) on top of In this section, we describe network functions that need to
commodity hardware. be automated by using machine learning techniques. These
functions are shown in Fig. 2 and described below.

– 75 –