Model-driven & AI-Enabled Inter-Cloud Optimization Architecture - PowerPoint PPT Presentation

Model-driven & AI-Enabled Inter-Cloud Optimization Architecture and Benefjts Ramki Krishnan

Introduction • What did we talk about so far? • Model-driven & AI-Enabled Inter-Cloud Optjmizatjon • 5G/Edge Computjng Use Cases – Dilip Krishnaswamy • Let us talk about the architectural requirements

End-to-end Reference Architecture – ONAP Perspective *** This diagram is discussion in progress and not fjnal *** Discussion in Progress: Edge Automatjon Through ONAP WG (htups://wiki.onap.org/display/DW/Edge+Automatjon+through+ONAP)

Architecture - What do we need? (1) • Centralized Resource Management/Optjmizatjon • 1000’s of Clouds • Probabilistjc Decisioning • Multjple Solutjon Choices – Aggregate Data for scale, Data Collectjon tjme lag etc. • Several Constraints, need fmexibility to easily add new constraints • Cost (Partner Cloud, Private Cloud etc.), Service SLA (Latency etc.) • Data Sources are ofuen Aggregates, examples below • Partner/Public Cloud -- Cloud Region & Tenant Resource (Compute/Network/Storage) Available Capacity & Utjlizatjon; Cloud Region Energy Utjlizatjon • Private Cloud – Above + Cluster Capacity/Utjlizatjon etc. • Policies are ofuen sofu constraints, examples below • Find Cloud Regions(s) with least resource/energy utjlizatjon, least cost etc. • Automatjon Intelligence (AI) through Machine Learning (ML) • Use ML (non-linear regression etc.) techniques on operatjonal data to predict the thresholds for sofu/hard constraints • Update the thresholds for sofu/hard constraints in a closed-loop operatjon Discussion in Progress: Edge Automatjon Through ONAP WG (htups://wiki.onap.org/display/DW/Edge+Automatjon+through+ONAP)

Architecture - What do we need? (2) • Edge Resource Management/Optjmizatjon • 1-10 Clouds • Accurate Decisioning • Single Solutjon Choice • Data Sources are Atomics, examples below • Partner/Public Cloud -- Workload (VM/Container) Resource (Compute/Network/Storage) Available Capacity & Utjlizatjon etc. • Private Cloud – Above + Host Capacity/Utjlizatjon etc. • Inter-cloud latency, bandwidth etc. • Policies are ofuen hard constraints, examples below • Find Cloud Regions(s) with SR-IOV support • Automatjon Intelligence (AI) through Machine Learning (ML) • Same as Central Resource Management/Optjmizatjon • Note: For some deployments, this functjon could be combined with the central component Discussion in Progress: Edge Automatjon Through ONAP WG (htups://wiki.onap.org/display/DW/Edge+Automatjon+through+ONAP)

Resource Management/Optimization and Related Components Architectural Framework Model-driven Optjmizatjon Libraries – Minizinc etc. Flexibility to add Custom Optjmizers ML Component Use Operatjonal data to predict the thresholds for sofu/hard constraints • Designer & Developer friendly Domain-Specifjc Modelling Language for Service Placement/Scheduling Policy • Address Central/Edge Resource Management/Optjmizatjon Requirements • Masks the Mathematjcal complexity of optjmizatjon algorithms through Modelling • Flexibility to add Custom optjmizers especially for Edge Resource Management/Optjmizatjon • Drive Service Creatjon Agility for 5G, Edge Computjng etc. Discussion in Progress : ONAP Optjmizatjon Framework (OOF) -- htups://wiki.onap.org/pages/viewpage.actjon?pageId=3247288 Note: This is an exemplary architectural framework/implementatjon choice

Upcoming Talks • “Recent Trends in Constraint Optjmizatjon and Satjsfactjon” -- Nina Narodytska • “SCOR: Sofuware-defjned Constraint Optjmal Routjng platgorm for SDN” – Siamak Layeghy • Model-driven Minizinc applicatjon for constrained-based Routjng