...
- Cloud agnostic placement attributes are targeted to abstract the following cloud specific placement attributes
- HPA attributes (e.g. Smart NIC Family, GPU Family) based on Cloud specific HW/SW support
- Normalized CPU capacity for VMs/Containers based on Cloud specific HW support
- Reference for CPU Normalization: https://d1.awsstatic.com/whitepapers/Demystifying_vCPUs.df200b766578b75009ad8d15c72e493d6408c68a.pdf
- Fine-grained Placement attributes based on Cloud specific SW support
- e.g., Rack-level Anti-affinity-> Azure: Fault-Domain, AWS: Placement-Group
- e.g., Exclusivity -> Azure: Isolated VM, AWS: Dedicated Host
- e.g., Fine grained QoS -> VMware Minimum guarantee, Kubernetes Burstable Class
- For the 5G Service/VNF placement example in Note 3
- Multi-Cloud interprets cloud agnostic intent as Physical DC Endpoint and translates to cloud-specific placement attribute such as Availability Zone
- For this example, each distributed physical DC is in a separate Availability Zone for a OpenStack-based Cloud
- Multi-Cloud interprets cloud agnostic intent as Physical DC Endpoint and translates to cloud-specific placement attribute such as Availability Zone
- Cloud Agnostic Placement Example
- Intent: Rack-level Anti-affinity for VMs within a VNF
- Different Realization for different clouds
- OpenStack-based (open source, Wind River, VMware VIO)
- Heat Template – policy-group, anti-affinity policy
- Azure
- Fault-domain
- AWS
- Placement group
- OpenStack-based (open source, Wind River, VMware VIO)
- Cloud Agnostic Networking Example
- Intent: High Performance Intra-DC data plane networking
- Several Realization Options per Cloud – the final realization choice could be based on least current resource usage or other criteria, cloud capability, cost etc.
- Overlay in SmartNIC
- Gateway in SmartNIC
- Overlay in DPDK-based switch/router
- Gateway in DPDK-based switch/router
- Overlay in ToR
- Gateway in a ToR
- Gateway in a HW appliance
- Realizations which are fixed
- Underlay maps to ToR/Network Fabric
- No CPU usage for data plane networking maps to VMs/Containers with SR-IOV support
- Cloud Agnostic Resource Slice Isolation (Tenant) Example
Intent: Resource Slice (Tenant) Isolation
Details in next section on "Cloud Resource Partitioning for Differentiated QoS"
- Several Realization Options per Cloud – the final realization choice could be based on least current resource usage, cloud capability, cost etc.
- Guaranteed Resource Slice (hard isolation) for various infra Resources (CPU/Memory/Network)
- Burstable Resource Slice (soft isolation) for various infra Resources
Note 8:
- Cloud Agnostic Intent Execution Workflow
- 1) OOF → MC runtime check for a specific cloud region which can support capability/capacity/cost metrics (enhance current capacity check API appropriately)
- 2) MC processes request from OOF
- Retrieve target Cloud Region specific policy from "Policy"
- Evaluate each of the cloud specific options in the policy from a perspective of resource allocation, utilization and cost
- Return to OOF the option which minimizes <resource allocation, utilization and cost> and the net value of the option
- In case the requirement cannot be met, return appropriate error
- Cache the option in MC for the specific cloud region as part of the VNF deployment workflow
- 3) OOF processes request from MC and picks the target cloud region which maximizes the net value
- 4) SO → MC VNF deployment request is processed by MC
- MC looks up the cache for the target cloud request with the VNF deployment request details
- MC replaces appropriate details in the cloud specific template (Heat, ARM etc.) based on the chosen option in the cache
- MC deploys workload on target cloud region with dynamically modified cloud specific template
- MC removes the cache entry for the specific cloud region for the specific VNF deployment
...