Page 66 - Kaleidoscope Academic Conference Proceedings 2022

2022 ITU Kaleidoscope Academic Conference

To enable adaptive behavior, and thus realize self-driving as well as the east-west bound interface for decentralized
networks, management and control planes require controller settings, have not reached adequate maturity levels.
decentralization so that relevant functions can execute very In a parallel effort, the P4 language [4] addresses a more
close to the infrastructure (if not in it) and base their general problem than OpenFlow and has been designed to
decisions on local knowledge. Such a setting would allow the program the behavior of the forwarding plane irrespective of
timely detection of events that adversely affect the service the supported network protocol and the type of controller.
performance, and the swift computation and enforcement of
appropriate remedy actions, for example rerouting through Although P4 has gained traction in the networking
an alternative path or selecting another server hosting an community, the language is fairly low level and therefore
application instance. In addition, to warrant the performance somewhat complex even for simple packet processing
of demanding services delivered over large network operations. We believe that simpler approaches are needed,
infrastructures, real-time telemetry is needed so that an which do not necessarily require expert knowledge on the
accurate view of the resources can be constantly maintained. underlying networking technologies. The abstraction level at
The vast amount of collected information can overwhelm which the resources are programmed should be raised.
network management systems, hence efficient mechanisms Instead of composing code that details ’how’ a goal is
are needed that can dynamically self-tune and achieve the achieved, directives are used to define ’what’ needs to be
right trade-off between accuracy and overhead [22], [23]. achieved. Alongside a declarative approach for expressing
the desired network behavior, a mechanism is needed that
2.2 High-level programmability automatically decomposes high-level goals to specific
configuration settings, thus allowing the network to flexibly
Network programmability has always been a challenging adapt to changes in business objectives, administrative goals,
issue with solutions traditionally involving complex low- and application requirements. Research on intent-based
level scripts and multiple vendor-specific interfaces, thus networking [6], [25] has taken some initial steps in this
making desired changes in the behavior of network devices direction, but a concrete language and decomposition
a daunting task. Programmability has been a topic of solutions have not been made available yet.
discussion since the late 90s with proposals on open APIs
[14] and active networks [3], [24]. While the idea of 3. CLOUD NATIVE NETWORKING
programming control information in headers that defines the
treatment of packets in the network initially seemed Cloud computing has brought enormous advantages in terms
promising, it was eventually dismissed on the grounds of of elasticity, scalability, and automation in computational
significant security risks. and storage infrastructure. However, large cloud providers
tend to be based around a handful of large data centers
In an effort to reduce the network administration complexity, strategically positioned at central locations on the Internet.
the policy-based management technology [21] was Users are, therefore, often located hundreds or thousands of
extensively researched around the same time, in which kilometers away from the computational nodes hosting the
policies govern the network behavior offering a certain services they are accessing. This implies network and
degree of programmability. These are technology- processing latencies in the order of hundreds up to thousands
independent rules that enhance the functionality of network of milliseconds for many users in addition to the
devices by introducing interpreted logic that can be inefficiencies and cost implications of transferring data over
dynamically changed without modifying the underlying large distances and long network paths in the case of data-
implementation. Alongside the core PBM technologies, intensive applications. This also has additional non-technical
techniques to refine high-level goals to concrete disadvantages such as data being stored and processed in
configuration parameters had been developed at the time [2], regions under different jurisdictions to those where the users
[16]. These, however, were not fully automated and were are located. In terms of security and privacy, the
also based on application-specific policy refinement patterns, concentration of clouds in a few central locations turns them
which limited their applicability to specific domains, e.g. into easier targets for (cyber and physical) attacks and
quality of service and security. Another issue concerned the facilitates surveillance taps into the fiber optic links
possibility of configuration inconsistencies as a result of interconnecting data centers.
policy conflicts [5].
The above issues drove research in academia and industry to
More recently, the decoupling of the control plane from the investigate distributed deployments of a vast array of
forwarding hardware in the Software-Defined Networking computational resources in the vicinity of both the users and
(SDN) paradigm [12] empowers operators with a highly data sources [18], [20]. Based on the flexibility concerning,
flexible approach to control the behavior of the network, for example, the locations at which applications execute and
since control functions/policies (realized in soft form) can be the selection of nodes from where user requests are served,
easily added and removed. While the southbound interface the edge compute paradigm envisions an environment where
of the SDN architecture, i.e. OpenFlow, has undergone services can be accessed with a much lower latency, while at
several iterations and allows us to define the treatment of the same time reducing the traffic footprint on the network
traffic in switches, the northbound interface that supports the and enabling better resilience to failures and attacks.
interaction between the controller and external applications,

– 20 –