The Role of Machine Learning in Network Automation Alberto Leon-Garcia University of Toronto alberto.leongarcia@utoronto.ca Acknowledgment to: Dr. Saeideh Parsaei Fard and Iman Tabrizian
Outline • Context: • Network Automation • Addressing global challenges • Networks & Application Platforms • MAPE-K Loops and ML Pipelines • AI as a Service • Challenges & Recent Work 2
AI and ML! Deep Learning: MLP, CNN, RNN Neural networks Machine learning algorithms • Supervised learning • Unsupervised learning • Reinforcement learning • Online learning AI engines: • Rule engines • Expert systems • Evolutionary algorithms source : www.atis.org. report 2018, Evolution to an Artificial Intelligence Enabled Network 3
Use Cases of AI in Networking https://www.etsi.org/images/files/ETSIWhitePapers/etsi_wp22_ENI_FINAL.pdf 4
Ridesharing With full penetration of ridesharing, the number of cars on the road could be reduced to 1/3. 5
EVs, Ridesharing, Renewable Energy • Real-time control of EV’s • Ridesharing • Recharging from Renewables • Global Policy • Carbon Footprint • Energy Efficiency • Environmental Impact • Productivity • Immense Potential to address Urbanization Challenges 6
Air Quality & Climate Change • In 25 years, there could be no air pollution in Santiago Chile 7
Smart is all about Data! • Real-time Situational Awareness • Continuous Monitoring & Data Collection • Learning and Intelligence • Analytics & Machine Learning • Visualization • Anomalies, Trends, Forecasting, KPI Analysis • Smart Applications Enablement • APIs provide real-time intelligence to Decision-making
Making the Network Agile & Smart Major functionalities for software-defined control • Identify the state of the network • Disseminate data (efficiently) to where it is consumed • Analyze data; understand network & service behaviour • Decide what changes are required • Apply changes to the network • MAPE-K Loop!
MAPE-K Loop for Autonomous Network Management Knowledge Step 2: Analysis Step 3: Planning Step 4: Execution Step 1: Monitoring Data Analyzing, Preparing a plan and Destination Source Data Execution and Gathering and Optimizing related parameters for Domain Preparation Implementation Domain Storage and Learning action 10
ETSI Experiential Networked Intelligence Intelligent Components Evolution to Network Intelligence Dynamic network conditions • More services & users • Better network telemetry • Cost-effective AI and ML • Evolution to Mgmt & Ops Intelligence Actuation Sensing Human decisions • Complex policy decitions • Complex manual operations • https://www.etsi.org/images/files/ETSIWhitePapers/etsi_wp22_ENI_FINAL.pdf
Outline • Context: • Network Automation • Addressing global challenges • Networks & Application Platforms • MAPE-K Loops and ML Pipelines • AI as a Service • Challenges & Recent Work 12
Network & Application Platform 3 rd Party … Custom Urban Congestion SaaS Portal KPIs Planning pricing Apps APIs Analytics Algorithmic BIaaS Engines Engines Publish/Subscribe Overlay Information-Centric Data Dissemination PaaS End-To-End, Multi Domain, Orchestration Monitoring & SDI Resource SDI Manager Topology Manager Management Analytics Software Defined Infrastructure Access/Things Multi-Tier Cloud Controllers (SD) Network Controllers Controllers Resources Phys. 13
Questions • How to deploy MAPE-K loops? • Where to deploy ML? • To share or not to share? • AI as a Service?
Outline • Context: • Network Automation • Addressing global challenges • Networks & Application Platforms • MAPE-K Loops and ML Pipelines • AI as a Service • Challenges & Recent Work 15
3GPP Network Data Analytics Function NWDAF Service-based architecture for 5G control plane: Components query NRF to discover • & communicate with each other Cloud-native • NWDAF Network Analytics logical function • Collection of data • Provide slice-specific network • analytics to other Network Functions Enabler of network automation •
ENI Assisting MANO Intelligent Components 17
ENI Assisting SDN Intelligent Components 18
ITU ML-Aware 5G Architecture ML Overlay • Logical entities (& functionalities) combined to form analytics fcn • Independent of underlying netwks • Common vocabulary & nomenclature for ML fcns & interfaces • Enables interoperability of ML apps with heterogeneous networks • Rapid provisioning of ML apps pre-processor policy • Cost-effective AI and ML Technology-specific realization distributor collector • Apply logical ML overlay to specific technoolgy • 3GPP, MEC, EdgeX, … https://www.itu.int/en/ITU-T/focusgroups/ml5g/Documents/ML5G-delievrables.pdf
Unified Architecture for ML in 5G Management Subsystem Orchestration • Management (VNFM, EMS) • Platform (VIM) • ML Function Orchestrator • Multi-level ML pipeline Overlaid on existing NFs • Instantiated by MLFO • Multilevel chaining • Closed-loop subsystem Allows ML pipeline to adapt to change • Sandbox w simulator & real data • https://www.itu.int/en/ITU-T/focusgroups/ml5g/Documents/ML5G-delievrables.pdf
Multi-level ML in MEC & 3GPP ML pipeline 3 → 6 • Local predictions at NMS affect configurations in different domains (e.g. optimizations). 9 → 2 → 4 → ML pipeline1 • inputs from RAN & UE/RAN to make predictions at CN (e.g., MPP). 10 → 7 → ML pipeline 2 → 8 • Inputs from MEC platform to make predictions at the edge and apply them to MEC. Could also use side information from the UE and RAN (e.g., caching decisions made at the MEC). 3 → 4 → ML pipeline1 → 5 • Inputs from CN and possibly UE/RAN inputs to make predictions at CN ,and apply to NMS parameters, that in turn affect configurations in different domains (e.g., SON decisions made at the CN). https://www.itu.int/en/ITU-T/focusgroups/ml5g/Documents/ML5G-delievrables.pdf
5G Slice Broker in NEC Challenges; How to map heterogeneous service • requirements onto the network resource availability? Solution: 5G Network Slice Broker A mediator should be • interposed between external tenants and mobile network management
5G Network Slice Broker • Resource monitoring: e.g., resource blocks, MCSs • Machine Learning operations for traffic forecasting (online reinf. learning ) • Admission Control for network slice requests (based on forecasting info) • Support for multiple classes of Network Slices SLAs • Heterogeneous QoS traffic requirements (data rate and latency)
Machine Learning is NOT Only About Model Code Hidden Technical Debt in Machine Learning Systems https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf
Applying Big Data & ML Model to SON/SDN Controller L. Le, D. Sinh, B. P. Lin and L. Tung, "Applying Big Data, Machine Learning, and SDN/NFV to 5G Traffic Clustering, Forecasting, and Management," 2018 4th IEEE 25 Conference on Network Softwarization and Workshops (NetSoft) , Montreal, QC, 2018, pp. 168-176.
Use cases: • Traffic forecasting L. Le, D. Sinh, B. P. Lin and L. Tung, "Applying Big Data, Machine Learning, and SDN/NFV to 5G Traffic • Congestion avoidance Clustering, Forecasting, and Management," 2018 4th IEEE Conference on Network Softwarization and Workshops (NetSoft) , Montreal, QC, 2018, pp. 168-176. • Abnormality detection doi: 10.1109/NETSOFT.2018.8460129 • Energy Saving 26
Acumos AI • Started by AT&T and Tech Mahindra • Currently a Linux Foundation project • With the goal of making it easier to build, share, and deploy AI apps • Acumos Marketplace
Outline • Context: • Network Automation • Addressing global challenges • Networks & Application Platforms • MAPE-K Loops and ML Pipelines • AI as a Service • Challenges & Recent Work 29
SDI based Architecture for AI-aaS CF3 Chain of network functions of Slice Application n CF1 CF2 CF5 Chain of network functions of Slice Health CF3 CF4 CF1 CF2 OTT CF5 Chain of network functions of Slice AR/VR T- CF4 Chain of network functions of Slice URRLC Networking Application Layer Business Application Layer or AI-aaS Application Intra Slice Over the top (OTT-Layer) T2 (NAL) Sandbox TO Drones, e-factory, robotics, haptic, image e.g., load balance, QoS assurance, Anomaly AI-aaS X-MKL Sandbox processing, Big data analysis Plane Orchestrator (s) detection, SLA management T1 MKL –chains of NAL MKL –chains of OTT AO- TNAL MM X-MKL- AO-SM Manager (s) Open InterfaceI-2 T3 MM-SM Computation domain Controllers Slice Controllers SDN Controllers Intra AI-NAL SDI- Sandbox Open Interface-1 manager(s) SDI Transport subdomain Core Internal Third Parties AI-aaS AI-aaS subdomain subdomain subdomain Management Plane Training Plane
Kubeflow • An open source project by Google • Make it easier to run ML workflows on Kubernetes • Open sourcing the way Google ran Tensorflow • Democratizing ML • Response to increasing number of ML users at Google • Active contribution by SAVI members to this project
Recommend
More recommend