contact contact
Tags: Innovation, Software and Automation
December 11, 2018

CenturyLink and Infinera: On the Path Toward the Cognitive Network

By Sharfuddin Syed, Distinguished Engineer, Infinera

and Jack Pugaczewski, Distinguished Architect, CenturyLink

A new generation of end-user applications portends to not only drive continued demand for increased capacity, but also to create formidable operational challenges resulting from rapid and unprecedented shifts in traffic patterns. Therefore, we hear so much in the industry today about the importance of network automation. As fiber is pushed deeper into the network and becomes increasingly critical to emerging services such as the Internet of Things (IoT), 5G and augmented and virtual reality (AR/VR), what will it take to transform the underlying transport infrastructure to meet these challenges? What innovative solutions are required to evolve the network beyond simple automation to become a truly cognitive network?

Drivers for network automation
Figure 1: Changes in traffic patterns and network complexity drive the need for network automation

The Cognitive Network

The cognitive network is a multi-layer, self-aware, self-organizing and self-healing infrastructure that can take predictive and/or prescriptive action based on real-time knowledge gleaned from collected data and experience. While no network can completely plan or run itself, the cognitive network will dramatically reduce the number of manual tasks required across a multi-layer, multi-domain, multi-vendor network. There are three main building blocks of a typical cognitive network:

  1. The streaming of time-stamped data, known as key performance indicators (KPIs). This data is used to detect any anomalies over time and essentially uses machine learning (ML) algorithms.
  2. Machine learning components comprised of understanding what happened and the reason why, learning inference and reporting the results.
  3. An artificial intelligence piece that enables the use of the inference and comes up with an automatic or policy-driven action to close the loop.
The building blocks of cognitive networking
Figure 2: The path to cognitive networking: the building blocks

Open, Software-driven Networking Intelligence

CenturyLink and Infinera recently joined forces to showcase innovation toward building a true cognitive network and to demonstrate data science use cases within the Metro Ethernet Forum (MEF) Lifecycle Service Orchestration (LSO) architecture.  Featured at the recent MEF18 and SC18 industry conferences, this proof of concept (PoC) marked a key step toward advancing open application programming interfaces (APIs) and microservices-based application development critical to the implementation of data science solutions in a new era of communications.

Using a combination of open protocols, software-defined technologies and advanced networking intelligence, this PoC provided an innovative framework for showcasing data science and policy-driven applications in an open network architecture. Key components that enabled the PoC included:

  • Policy engine/API
  • Software-defined networking (SDN) controller and orchestration
  • Network elements with gRPC/gNMI agents
  • gRPC telemetry collector
  • Kafka-based streaming
  • AI/ML intelligent engine
  • Open standard APIs (MEF 60)
  • Message Queuing Telemetry Transport (MQTT) publisher-subscriber for notification application

Artificial Intelligence, Machine Learning and the Cognitive Network Future

This CenturyLink and Infinera PoC demonstration featured automatic reconfiguration of the network (e.g., dynamic Ethernet Virtual Connection creation) based on intelligent decision criteria, which was supported by well-defined and configured policy engine rules and supplemented with time-series telemetry information. It also highlighted the integration of various third-party toolkits that can be easily integrated into the MEF LSO ecosystem in alignment with MEF’s Third Network vision, resulting in practical realization of the MEF standards work.

The PoC will help advance standards activity in the areas of artificial intelligence and machine learning while integrating the MEF LSO APIs toward intelligence-based automation, which represents one of the key building blocks of the cognitive network. The PoC also illustrated a means to advance the user experience in the DevOps area by promoting more open APIs, ensuring the fast and agile introduction of features and using a generic policy engine and actions.

CenturyLink and Infinera proof of concept implementation
Figure 3: CenturyLink and Infinera PoC implementation of data science applications for the MEF LSO architecture

Dynamic Additions of User-defined Use Cases Based on Flexible Policy Management

The architecture used in the POC is extensible. Additional use cases, including but not limited to those described below, can be easily added depending on the network deployment scenarios. The policies can be changed in real time as required by the network administrators.

  • Predictive cognitive analytics, wherein a failure can be predicted across the entire network well in advance. Preventive maintenance can thus be initiated before a complete failure occurs in the network.
  • Automatic reconfiguration of network services (addition, deletion or updates) can generate notification alerts to the operations and business support systems for tracking service level agreements (SLAs) and/or billing systems.
  • Security anomalies detection (using traffic pattern anomalies, power level anomalies and deep packet inspection).
  • Policy-based calendaring and energy-efficient networks by means of automatic power savings.
  • Mitigation of linear and non-linear impairments and anomalies in optical networks using machine learning.
  • Multi-layer fault correlation and SLA monitoring.

Key Takeaways

While embarking on this PoC, the following key areas of importance were identified for the advancement and adoption of AI/ML-based cognitive networks:

  1. Data is the cornerstone. It is extremely critical to stream the “value-added” KPI data at a decent frequency interval with time stamping. The streaming of KPI data needs to be based on flexibility (policy-based) such that the KPIs being watched will result in deterministic inference in the learning process with very high confidence levels. For multi-layer, multi-vendor deployments, KPIs need to be standardized at all layers of the network. This ensures that all operators (service providers and customers) can adopt the same methodologies and terminology for the KPIs.
  2. A fine mix of machine learning based on the large data sample, together with intelligent domain knowledge expertise, results in solving critical use cases. The AI/ML provides two functions. The ML part uses data samples to arrive at an inference. The AI part uses the inference and domain knowledge (intelligence) to accomplish a given use case by orchestrating the network configuration directly or via the policy-based action. In either case, a closed loop is achieved without human intervention. Please note that this architecture can also encompass the basic use cases in which closed loop automation is achieved by simply monitoring any attribute of the sensor device. In such scenarios, the ML part can be skipped while the AI part still provides the required intelligent piece of functionality.
  3. Performance, reliability, scalability and agility. The AI/ML intelligent engine can be installed in the cloud. Depending on the number of sensors being monitored from many network elements, the central and graphics processing units, memory and data store can be adjusted to support performance and scale. The high availability requirement is also met by deploying the AI/ML in multi-core redundant instances. It is important to avoid false positives while inference is made. Proper hysteresis is built into the ML algorithms to avoid transient inferences. The robust handling of time-series KPIs so that they arrive at reliable inferences is key to a highly dependable cognitive-based automated network. It is also apparent that the amount of usable open source solutions can be easily leveraged to build applications at a faster pace.
  4. The customer can customize their service performance using open policy API(s) and open microservices.

Pushing the Boundaries Through Industry Collaboration

Cognitive networking is the result of seamless and highly dynamic interaction between software and hardware assets across network layers and brings networking to a new level of scalability, flexibility and automation. The path toward a true cognitive network will pave the way for innovations in service delivery that will redefine next-generation communications services for a wide variety of users and verticals, including banking and finance, government and education.

Many thanks to the CenturyLink and Infinera PoC team!

Click here to learn more about the recent CenturyLink and Infinera technology collaboration.

Legal Disclaimer