AI Driven Orchestration, Challenges & Opportunities Openstack Summit 2018 Sana Tariq (Ph.D.) – TELUS Communication
Agenda Service Orchestration Journey Service Orchestration Operational Challenges Closed Loop Orchestration and Dynamic Policy AI/ML Driven Orchestration
NFV and Orchestration Journey… 5G IoT customers defined services through customized user portals AI driven/managed AI driven/managed operations, capacity services (advanced) and applications Automated Assurance Applications Onboarding PNFs Building NFV Cloud Q2 2018 2016 2017 2018 2019 2020 2021 2022 IoT and OTT Services/ User-defined Services/International customers Increased Maturity OSS/BSS interlock, evolution of customer Portals, inventory compliance Increased Maturity of Catalogs/Templates/Blueprints to deliver software defined services
NFV Telco Cloud - is different … Self serve Reliability Automation APIs NFV Telco Cloud Scalability Fast and agile Standards based High Throughput Low Latency 24x7 Secure Cost efficient availability
NFV and Orchestration Journey… • Building robust cloud infrastructure Ticketing Alarms Billing Analytics • Virtualizing Applications End to End Orchestrator OSS/BSS • Provisioning workflows • Assurance EMS 1 EMS 2 EMS n • Automated scaling EMS DC WAN • Traffic steered through SDN network VNFMs VNFMs Service Assurance • Monitoring through holistic service VNF 1 VNF 2 VNF 3 VNF n SDN assurance Controllers VNFs • Integration with OSS/BSS VIM • E2E Service Orchestration plays are NFVI major “orchestration” role WAN
Orchestration: Commercial or Open Source Commercial Opensource Vendor lock-in/proprietary plugins Vendor agnostic/shared community plugins Higher licensing cost Significantly lower costs Trusted support model Self/Community support R&D driven roadmaps Community driver roadmaps Depends on Company size Depends on community participation Opensource SI Vendors
Orchestration Functional Elements
Cloud Robust Cloud supporting automation features, APIs, elasticity etc. SDN Network Complete softwarization of DC and WAN for on-demand creation of services ORCHESTRATION Analytics SUCCESS SERVICE Robust analytics cross-functional domains for scaling, healing & optimization for cloud and services DevOps Effective DevOps culture and support for short time-to-market and efficiency targets Data Models Homogeneity across stacks for consistent data models for seamless integration, reusability and abstraction
Chaos of Multi-Vendor Multi- Domain… NaaS/OTT Voice OTT 5G IoT VoD Customer facing services vEPC S-GW P-GW MME ... vIMS E2E Service S-CSCF Orchestration SBC ... TAS I-CSCF SD-WAN Enterprise NFV Cloud
Software Defined Service Operational Challenges… 1 Service Design 6 Debug Troubleshooting 2 Orchestration (testing in a sandbox) 4 Orchestration (runtime) Launch 3 Service Package and failure distribution 5 Assurance Offline in lab Services Operation in real-time
Closed Loop Orchestration Static… E2E Service Orchestration Service Orchestration BigData/Hadoop Policy Engine VNFM Analytics LB vFW vFW VAS Collector
Closed Loop Orchestration AI/ML… BigData/ Hadoop Predictive/proactive Inputs Analytics Service Orchestrator Life Cycle Service Definition Testing & Workloads Management Workflows Validation Onboarding Resource APIs Adaptors SDK Policy Inventory Orchestrator Reactive Inputs Conditional Rules Mapped Actions Cloud Services Security Network
AI/ML Orchestration Data Surge Static Policies Dynamic Policy Analytics Model HOW? WHY? • 5G Services/ IoT • Reactive decisions • Strategy to develop • Robust Analytics increase in traffic • False Spikes dynamic policies and framework for volume and patterns • Too many policies testing feeding AI/ML & conflicts systems Customer Cloud Network Security Experience Optimization Optimization What? Closer to the edge • • Energy optimization • Traffic optimization SDN • Proactive/predictive threat Service Healing • • Capacity optimization controller identification Service optimization • • Faults prediction & healing • QoS based routing • Closed loop decisions to Differentiated QoS • • Fast troubleshooting attacks mitigation
Building an AI Application STEP 1 STEP 3 STEP 5 RESULT Testing phase: Problem Prepping the Experiment with definition data: cleansing, Models, Pick best normalization, Models and feature create feedback engineering Loop ? STEP 2 STEP 4 STEP 6 Gather real, Training phase: Evaluation, sizeable training robust training predictor data environment improvement START and re-training PROCESS
AI is becoming easier… Add your own description here. Add instructions, or additional information about this slide here. You can delete this item or any item on the slide. Acumos Raw AI/ML ML Libraries AI Projects Data Scientists Python libraries – H2O, TensorFlow Predictors & needed, higher lower complexity Models developed complexity ONAP leverages by shared SPARK, Mlib, community effort + MALLET, WEKA data etc. Computing Power Accessible Data • Better Analytics Engines, • Low Cost • Better data • High Power organization • GPU, NPU • Large volumes of availability managed Data
ONAP Intelligent Closed Loop Architecture… Design Time Run-time DCAE POLICY EXECUTION ENGINE CLAMP (Closed Loop Automation AI/ML and Management Platform) Policy Execution (Python GUI that generates TOSCA Policy1 DMAAP Policy1 Library) Synthetic Data Service Orchestration (Extension) Hadoop SDC (Service Design Creation) Ceilometer collector SDN-Controller APP-C (GVNFM) Condition Cloud Region1 CPU Utilization >80% Action Shift VNF/ Ceilometer workload from Cloud Region 1 to Ceilometer Cloud Region2 VNF1 VNF2 TOSCA Policy1 TOSCA Policy1 Cloud Region 1 Cloud Region 2
ONAP Policy Execution Flow… 5 2 DCAE DCAE POLICY EXECUTION ENGINE 11 10 CLAMP 9 Analytics 12 Policy MS AI MS 13 8 3 DMAAP Service Orchestration Hadoop Cluster 1 7 14 4 SDC Ceilometer Collector SDN- APP-C Controller (GVNFM) 15 6 Ceilometer Cloud Region 1 Cloud Region 2
AI/ML Orchestration Industry Verticals Customer with Customer with Customer with 5G Customer with IoT real-on-demand Customer with off- real-time Voice Services: real time, Traffic viewing line downloading Traffic high bandwidth Study of VNFs performance requirements (Dependency: Dependency: Dependency: Dependency: Dependency: Closer to Edge Performance/ TTR/ Throughput only (MME, S/P-GW, vSBCs etc.) Performance/ TTR Performance/ TTR Location) throughput BigData/ Customer/Resource Facing Portal Inventory Hadoop/ Analytics NSD and VNFD Package description to associate color tags for VNF types API Mediation Policy Rules to support these use-cases Service Orchestrator · Initial Service Definition Life Cycle Testing & Workloads Placement Management Workflows Validation Onboarding · Cloud Optimization related DCAE: Dynamic Policy and Cloud actions · Customer QoS optimization use-cases Resource APIs Adaptors SDK Policy related Inventory Orchestrator actions Develop ML Models for cloud NFV Cloud optimization use-cases Enhance ceilometer & support Zone1 PODs Zone2 PODs Zone3 PODs Edge POD: High cloud optimization Optimized for Optimized for Low Optimized for Low Bandwidth, Closer to configurations Efficiency/ Latency/HTTR (real- Latency/HTTR (real- customer throughput time traffic) time traffic)
Sana Tariq (Ph. D.) Sana.tariq@telus.com
Recommend
More recommend
