An ambitious programme geared to create a radically new architecture for the UK’s internet and telecommunications infrastructure
A series of short technical talks from PhD students and Researchers on the project..
Unikraft: Fast, Specialized Unikernels the Easy Way
PhD Student Lancaster University
30th October 2021, from 13.00 - 13.30
The increasing softwarization of network technologies has mold computation as an essential resource to deliver next-generation network services. Nonetheless, the adoption of cloud technologies in order to manage and virtualize compute resources in the network infrastructure, limits the ability of operators to improve the resource efficiency and responsiveness of network services. In this talk, we will present a novel toolchain and library operating system for the cloud; Unikraft. Unikraft builds highly specialized unikernels, software bundles that consist of a target application along with just the operating system primitives and libraries features it needs to run. The Unikraft platform supports a wide range of VNF appliances (e.g. Click programs, HAproxy) and Unikraft kernels can run with minimal resource requirements (image size ~ 1Mb, memory footprint ~ 10 Mb) and fast boot times (30-40 msec). Furthermore, we will elaborate on the benefits of OS configuration tuning on cloud application performance and present Wayfinder, an holistic OS measurement platform which allows users to explore the configuration space of modern OSes and run custom OS benchmarking scenario in a reproducible way.
Alex is a 3rd year PhD student at Lancaster University focused primarily on the optimization of serverless applications and lightweight Virtual Network Functions (VNFs) constructed using specialized library Operating Systems known as Unikernels.
Bad Apples and Good Labels: Learning in Real-time Fault Detection
Lecturer in Statistics, Lancaster University
18th January 2022, from 13.00 - 13.30
When monitoring telecommunications networks, it can be a relatively low-cost activity to flag anomalous patterns in data streams. It can be much more costly to ask engineers to respond to or counterfactually label these anomalies, and there is often is a high proportion of patterns which are anomalous in a statistical sense, but innocuous in an operational sense. This talk describes an approach we have developed which can work as a mediator between anomaly detection algorithms and experts. The approach, based on the so-called ‘apple tasting model’, learns (in an online fashion, from limited expert feedback) to escalate only operationally-relevant anomalies to the expert, further automating the process of fault detection and monitoring.
James is a lecturer in Statistics at Lancaster University and a researcher on the NG-CDI project. His research considers the application of statistical thinking to sequential decision-making, particularly in settings with complex data structures. He has research interests in multi-armed bandits, recommender systems, and time series.
Advanced Topics in Deep Learning
Research Associate, University of Bristol
8th February 2022, from 13.00 - 13.30
The last decades have witnessed rapid progress in Artificial Intelligence (AI), Machine Learning (ML) and Deep Learning (DL) technologies. As a subset of ML, DL is essentially using neural networks with multiple layers to solve ML problems. Through learning from large amounts of data, DL automates feature extraction, enabling predictions with incredible accuracy. In this talk, the basics of machine learning and deep learning will be introduced, followed by more details around advanced DL topics including deep reinforcement learning (DRL), self-supervised learning, graph neural networks and several specific applications in telecommunications.
Dr Xiaoyang Wang is a Research Associate in the Communication Systems and Networks Group, Department of Electrical and Electronic Engineering, University of Bristol. Her current research focuses on machine learning, especially reinforcement learning, and its applications in next-generation network management.
Jonathan Thomas is a PhD student in the Communication Systems and Networks Group, Department of Electrical and Electronic Engineering, University of Bristol. His research focuses on the application of cooperative Multi-Agent Reinforcement Learning within next-generation communications networks.
The value of information for dynamic decentralised criticality computation
PhD student, University of Cambridge
22nd February 2022, from 13.00 - 13.30
Smart manufacturing uses advanced data-driven solutions to maximise performance and resilience of daily operations. It requires large amounts of data delivered quickly. Datatransfer is enabled by telecom networks and network elements such as routers or switches. Disruptions can render a network inoperable, and advanced responsiveness to network usage is required to avoid them. This may be achieved by embedding autonomy into the network, providing fast and scalable algorithms that use key metrics for prioritising the management of a potential disruption, such as the impact of a failure in a network element on system functions. Centralised approaches are insufficient for this as they require time to transmit data to the controller, by which time it may have become irrelevant. Decentralised and information bounded measurements solve this by situating computational agents near the data source. We propose a method to assess the value of the amount of information for calculating decentralised criticality metrics. The method introduces an agent-based model that assigns a data collection agent to every network element and computes relevant indicators of the impact of a failure in a decentralised way.
aniv is a 2nd year PhD student at the Institute for Manufacturing, in the Engineering Department within the University of Cambridge. His research has evolved into studying the effects of varying the visible range around nodes for distributed calculation of criticality, with mind to find the sufficient threshold that allows the network to preserve criticality
Rethinking DevOps for Networks: Enabling automation via ultra-lightweight µVNFs
PhD student, Lancaster University
8th March, from 13.00 - 13.30
Modern network infrastructure is ever becoming more of a complicated web of overlapping technologies, APIs, and standards. Each evolution does not replace the old, but infrastructures are expected to cohabit all, old and new. Any fight against this reality simply becomes just another evolution to add. Service monitoring is a prominent victim of this, often needing many layers of abstraction to group together various interaction methods and APIs. And although these abstraction layers are becoming more proficient, they typically still need a layer containing bespoke implementations for each specific target.
This talk discusses a generalised approach to monitoring network services, hardware or software, UniProbe. Embracing the evolution strategy of networks, UniProbe leverages the benefits of unikernels to deploy highly-efficient targeted monitoring probes within and between network function chains directly from the MANO layer.
Will is a 3rd year PhD student in the network research group at Lancaster University. His research investigates the challenges unikernels face when being integrated into an ever more heterogenous infrastructure.
AI-Driven QoS Assurance for 5G Network Slices
PhD Student, University of Surrey
22nd March, from 13.00 - 13.30
Network slicing enables 5G systems to support various business verticals, e.g. eMBB, mMTC and uRLLC, with diverse and stringent requirements by creating multiple logically isolated networks on a common infrastructure. It is transforming telecommunication systems from “one size fits all” networks to service tailored networks provisioned on demand. Moreover, it dramatically reduces the required time-to-market of new services and empowers network operators to rapidly increase their pace of innovation and quickly deploy their in-house developed software to offer new services or improve existing ones, opening up new avenues of creativity in building efficient and robust communication networks and generating new revenues. However, network slicing may significantly increase network management complexity. Thus, Artificial Intelligence (AI) based network management automation is essential for network slicing implementation to reduce CAPEX and OPEX of the network.
In this presentation, I will present my work on deep reinforcement learning (DRL) based adaptive resource allocation methods for core and transport network slicing. In the first part of my talk, I will discuss a DRL-based auto-scaling method for 5G core network slicing. This method adaptively adds/removes user plane function (UPF) instances in each core network slice to meet the packet delay requirement of each slice while avoiding resource over-provisioning. In the second part, I will discuss my ongoing work on the development of a DRL-based smart packet scheduler for transport network slicing to guarantee the QoS of each transport network slice.
Abdirazak received his M.Eng degree in Electronics and Telecommunications Engineering from Universiti Teknologi Malaysia, Johor, Malaysia in 2014, and his B.Sc (Hons) degree in Electrical and Electronics Engineering from Omdurman Islamic University, Omdurman, Sudan in 2012. He is currently pursuing the PhD degree with the Institute for Communication Systems (ICS), University of Surrey, UK.