Re-arranging content...and, cleaning up....
General Background
A broad set of transformations are taking place:
- Business transformation: OTT services, faster TTM, Monetization
- Technical transformation: QoE, ULL, SDN/NFV/OMEC integration, Edge Analytics, Big data, Virtualization, Automation, C->E, R->E
- Architectural transformation: 4 views “NORMA-like” Cloud, ECOMP, Flexible architecture (RAN, Core, CDN, Application delivery, Automation, IoT, fog,..)
- Industrial transformation: ICT&E
To efficiently and effectively deploy 5G network supporting ultra low latency and high bandwidth mobile network, we need to deploy variety of applications and workload at the edge and close to the mobile end user devices (UE or IoT). That would include various virtualized RAN and core network elements, content (video), various applications (AR / VR, industrial automation, connected cars, etc.). We might deploy near-real time network optimization, customer experience / UE performance enhancement applications at edge. Edge cloud must support deployment of third party application (e.g. Value added optional services, Marketing, Advertising, etc.). We must deploy mechanisms to collect real time radio network information, process them in real-time (e.g. Geo Location data), summarize, anonymize, etc. and make them available to third party applications deployed at the edge or central location or outside service provider environment. Edge data collection could also be used for training machine learning models and fully trained models can be deployed at the edge to support network optimization.
The need
End users and other devices, cyber-physical systems will benefit from a broad set of context information that can enhance and enrich the delivery of a broad set of applications.
Service Deployment Goal
Deliver Application SLAs while minimizing TCO.
Edge Application Profiles
No | Application Classification (based on required RTT) | Application Examples | Network / Service Behavior Type | Deployment Component/ APIs | ONAP Managed | Edge Deployment Hard /Soft Constraint (Based on RTT) | Potential Application Provider | Casablanca Candidate | Additional Information |
---|---|---|---|---|---|---|---|---|---|
1 | Real-time (20ms -100ms) | In service path optimization applications which run in open CU-CP platform (also known as RAN Intelligent Controller, or SD-RAN controller). | Real-Time Network State Control | Open 5G CU-CP (CU - Control Plane) – VNFC. | Yes | Hard | NF Vendor/Service Provider/3rd Party | Yes | These applications include load balancing, link set-up, policies for L1-3 functions, admission control and leverage standard interface defined by oRAN / xRAN between network information base (or context database) and third party applications. Data collection through is B1 and implemented using x technology. |
2 | Near-real-time (500ms and above) | Slice monitoring, performance analysis, fault analysis, root cause analysis, SON applications, Optimization (SON Drive Test Minimization etc.), ML methodologies for various apps. | Network Analytics & Optimization | DCAE | Yes | Soft | NF Vendor/Service Provider | Yes | |
3 | Near-real-time (500ms and above) | Video Analytics, Video Optimization, Customer geoLocation information, Anonymized customer data etc. | Workload Analytics, Optimization & Context processing | Cloud Edge or Cloud Central | No | Soft | 3rd Party | NA. Out of scope for ONAP | The apps are OTT and the service provider is offering their infrastrcture as a service to OTT providers. |
4 | Real-time (10-20 ms) | Third party applications that directly interacts with the UEs, like AR/VR, factory automation, drone control, etc. | Workload Automation / AR-VR / Content, etc. | UE or Cloud Edge | No | Hard | 3rd Party | NA. Out of scope for ONAP. | These are third party applications, developed by enterprise customers (e.g. factory automation) or content creators (AR/VR applications). In this case, messages or requests or measurements directly go from UE (via UPF or GWs) to the applications and applications respond back. |
5 | same as 3) | same as 3) | Value Added Services + same as 3) | same as 3) + MEC/Cloud APIs (Note 1) | Yes | same as 3) | same as 3) | Stretch | Service Provider could be oferring video surveillance (video analytics/optimization apps etc.) as a service to enterprises. |
6 | same as 4) | same as 4) | Value Added Services + same as 4) | same as 4) + MEC/Cloud APIs (Note 1) | Yes | same as 4) | same as 4) | Stretch | Service Provider could be oferring factory automation as a service to enterprises. |
Note 1: API Details
- e.g., MEC APIs - Location info, Radio control info etc.
- e.g., Cloud APIs - IaaS/PaaS + Context Awareness (time, places, activity, weather etc.)
Edge Infrastructure Profiles
( example based on Akraino Edge Stack..but, need to generalize)
Profiles | Workloads | Compute | Networking | Storage | Control | Security | Edge Application Infrastructure |
---|---|---|---|---|---|---|---|
Large | Support for VMs and containers. Commentary:
| >50 Compute Servers Accelerators: SRIOV based QAT for Crypto and Compression acceleration. ML/DL Accelerators Compute profiles: Fixed number of profiles are expected to be supported. (Will add profiles) | SRIOV Networking for High performnace Data plane VNFs. vSwitch (OVS-DPDK) based networking for all other workloads Multiple leaf switches and two spine switches WAN - Underlay :
Underlay realization options
Overlay realization options
IPv4 and IPv6 support NAT44 with LSN (Large Scale NAT) support by providers. Support for dedicated public IP addresses Commentary: Network sharing among container and VM workloads will need to be supported. DVR (Distributed Virtual Routing) for forwarding packets locally among vSwitch based networks. Leaf/Spine switches for forwarding traffic among SRIOV based networks and for networks between vswtich and SRIOV based networks. Few fixed profiles for following:
| Block device support using Ceph Dedicated nodes for storage ( 3 nodes ) Storage profiles representing whether the nodes are dedicated for storage, use compute nodes for storage, Number of nodes for storage etc... Is support for Object storage required in Edges? | Dedicated nodes for control stack Automation Offload Platform (Offloading ONAP) at the Edge. Few control profiles
Automation Offload Platform profiles consists of following:
| Transport : TLS 1.2 and above between ONAP and Edge Services Infra Security: TPM 2.0/SGX for private key security and secret/password protection, Remote attestation to detect any software tampering of compute, storage and control nodes. | MEC Platform as a VNF to provide contextual information to Edge applications. |
Medium | Same as above | Same as above. Number of compute nodes are >10 and < 50 | Same as above | Same as above, except that there is no dedication of nodes to Ceph cluster | Same as above with respect to control, but Automation Offload Platform is not part of the Edge. No dedicated control nodes. Control functionality is shared with compute nodes. Support for K8S profile as it can support both VMs and containers | Same as above | Same as above |
Small | Same as above, but may support very less number of tenants | Same as above. Number of compute nodes are < 10 | Same as above, but no PE and CE at the Edge. Fabric itself acts as CE. | Same as above, no dedication of nodes to Ceph cluster | No control at the Edge No Automation offload platform at the Edge Regional sites are expected to provide control and AOP services. Support for K8S based control. | Same as above | Same as above |
Edge Infrastructure Profile Summary
- Distributed
- 1000's of edge locations of varying capacity
- Casablanca - Implementation
- 10-100 edge locations (simple starting point)
- Peformance-awareness
- GPU, FPGAs, SR-IOV etc.
- Casablanca - Implementation
- SR-IOV desired for Data Plane (5G CU-UP)
- NIC offload desired for tunnel encap/decap e.g. 5G CU-UP GTP tunnel
- Resource Isolation through fine-grained QoS
- Support both Latency-sensitive and General purpose applications
- Support ONAP Management plane components in the same cloud with Workloads
- Casablanca - Implementation
- Min/Max resource reservation model desired
- Security
- Workloads are often deployed in external (non-dc-type) locations and need HW security (TPM etc.)
- 3rd party applications which need additional HW security (VM, Containers in VM etc.) and SW security (Inter-component TLS etc.)
- Casablanca - Implementation
- Edge Clouds with private IP addresses, i.e. reachable via private connections
- For example, edge cloud in a public cloud provider reachable via AWS direct connect or Azure express route or Google partner interconnect
- Capacity constraints
- Very small footprint (few nodes per physical location), Medium footprint (10's of nodes per physical location), Large footprint (100's of nodes per physical location)
- Casablanca - Deployment
- Need number of cores per servers; Need storage capacity/pool
- Cloud Diversity
Private and Public Cloud Providers
- Casablanca - Implementation
- Note: ONAP currently supports private edge clouds based on VMware VIO, Wind River Titanium Cloud, Upstream OpenStack
- Desire to have at least one Public Cloud Provider (Azure, AWS, GCE etc.) as an Edge Cloud Provider
- ONAP central instantiates an Edge Cloud instance (blue cloud provider in gliffy) via a IaaS API to cloud provider
- ONAP central instantiates one or more ONAP edge components as need, e.g. DCAE
- ONAP central instantiates one or more NFs, e.g. 5G CU-UP/CP
- Configuration Diversity
- 5G Factory Automation, 5G General Mobility Services etc. – User Plane components (DU, CU-UP, UPF etc.)
Key Challenges w/Centralized ONAP Architecture for Network Function Edge App/Infra Profiles
WAN network bandwidth & latency issues for the following key functions
- Conveying Application & Infrastructure metrics, faults in real-time from Edge Cloud to Centralized ONAP for closed-loop fault management for a single edge site with a single Cloud control plane
- Conveying Application & Infrastructure metrics, faults, alerts in real-time from Edge Cloud to Centralized ONAP for dynamic network function closed-loop fault management/placement optimization across multiple edge sites with a single Cloud control plane
- Conveying Application & Infrastructure metrics, faults, alerts in real-time from Edge Cloud to Centralized ONAP for dynamic network function closed-loop fault management/placement optimization across multiple edge sites with multiple Cloud control planes
Exemplary Network Function Placement/Service Assurance Policy for demonstrating the aforementioned challenges
- "5G CU-UP VNF location to be proximal to a specific physical DC based on 5G DU, bounded by a max wire latency from physical DC"
Landscape for addressing aforementioned Challenges
- Cloud Providers
- Public Clouds such as Azure, Amazon, GCP do not support dynamic network function closed-loop fault management/placement optimization across multiple edge sites with single/multiple Cloud control planes
- Akraino Open Source
- Defines Blueprints for Edge Cloud using OpenStacks (VMs) and K8S (Containers)
- Plans to use ONAP for VNF/Service Orchestration for addressing aforementioned challenges
Hierarchical (ONAP Central, Edge) Architecture
Single Provider - Hierarchical Architecture
- Cloud Provider Business Unit:
- Provides hosting of Workloads, ie., IaaS/PaaS
- SP installs and manages ONAP in separate 'Management Cloud' instances
- SP installs and manages Network Services + 3rd Party Apps in separate 'Services/Apps Cloud' instances
- Provides hosting of Workloads, ie., IaaS/PaaS
- Cloud Provider Business Unit:
- Provides SaaS, eg., Analytics/Closed Loop as a Service, LCM of Apps, etc.
- ONAP Edge may not be needed
- Cloud Provider Business Unit:
- Types of virtualized cloud resource tenant and their characteristics
- Virtualized Network Workload Cloud Resource Tenant Category
- Network Management Cloud Resource Tenant Category
- Virtualized Application Workload Cloud Resource Tenant Category
- Application Management Cloud Resource Tenant Category
- Physical Network Function and their characteristics
- Part of Edge Cloud Orchestrator
- Types of virtualized cloud resource tenant and their characteristics
- Immediate interest to ONAP for Network Function use cases
- Virtualized Network Workload Cloud Resource Tenant Category
- Guaranteed
- Burstable (with minimum guarantee)
- Best Effort
- Network Management Cloud Resource Tenant Category
- Burstable (with minimum guarantee)
- Virtualized Network Workload Cloud Resource Tenant Category
Single Provider - Edge Functional Decomposition
Function | Statefulness | ONAP Project Mapping | Details |
---|---|---|---|
Inventory | yes | A&AI | |
IP Address Management (IPAM) | yes | SDN-C | |
Multi-Cloud Support | no | Multi-VIM | |
Initial Placement | no | OOF | |
Closed Loop Controller | no | APP-C | |
Closed Loop Policy | no | Policy | |
Infra Closed Loop Analytics | no | Multi-VIM, DCAE | |
App Closed Loop Analytics | no | DCAE | |
Logging | no | ||
Infra/App Monitoring events, statistics | no |
Single Provider - Sequence Diagram
ONAP Activity Goal #1: ONAP requires IaaS/PaaS attributes (see ongoing work – Edge Cloud Infrastructure Enablement in ONAP, 5G Items for Casablanca) from Cloud providers for Infrastructure profiles that allow Distributed, Highly-secure, Config/Cloud-diverse, Capacity-constrained and Peformance/Isolation-aware
ONAP Activity Goal #2: Define hierarchical ONAP Central, Edge Architecture/functional interactions (API reference points) to support aforementioned Application/Infrastrcuture profile in Any "Cloud" (internal Business Unit or external Partner) at Any "Location" edge, regional or central.
24 Comments
Ulas Kozat
The picture was discussed in the meeting. The separation of "access network" vs. "last mile network" does not make much sense to me as access networks provide the last mile solution. Maybe more inline with the picture is to use "customer-premise network". Even in that case why would an Edge Cloud in a customer premise have larger latency than an Edge Cloud in an access network?
Raghu Ranganathan
last mile vs access network: yes, the terms are confusing. maybe think last mile = internal LAN/campus network/tall-shiny-building. As example, inside a home. most of delay is due to slower speed links and/or best effort queuing in a wifi network, etc.
delay mentioned depends on location of measurement points (MEPs, MDs using IEEE CFM terms). As example, last mile network = host-to-CPE, whereas, access network = CPE-to-CO/metro
Arash Hekmat
The main work here is to define how a "Central ONAP" instance would communicate on its "Southbound" to/from "Edge Controller" instances. Where, "Edge Controller" instances could be of any type of: ONAP itself, other open source controllers (ONF CORD, Akraino, ...), or proprietary vendor controllers (Nokia, Huawei, ...).
The "Central Controller ← to/from → Edge Controller" communication is required in both directions: Central Controller → down-to → Edge Controller AND Central Controller ← up-from ← Edge Controller.
The main architectural work here is to define how ONAP's main "Southbound" components (Multi-Cloud, SDN-x, DCAE) would communicate both from a Central Controller to an Edge Controller (top-down) AND from an Edge Controller to a Central Controller (bottom-up).
This means Multi-Cloud, SDN-x, & DCAE need to act as both clients and servers. Today, on the "Southbound", Multi-Cloud and SDN-x only act as clients and DCAE only as a server. What is new here is that they all need to act as both clients and servers and expose a set of APIs for "Central-Edge" communications.
Another option is to create a "new" ONAP component that terminates all "Central-Edge" communications (APIs) and forwards them internally to Multi-Cloud, SDN-x, or DCAE.
ramki krishnan
Good points Arash. Please take a look at the notes below the gliffy diagram – we also need to cover the cases where Edge ONAP is optional and also Edge ONAP has only operational components (DCAE etc.).
Michael O'Brien
We have standardized on Kubernetes not Docker Swarm currently. We can collaborate the Multi-vim team around federated support for deployment and management of the clusters via helm. The current CD system runs on a single 5 node cluster on AWS - I would like to extend this to a Multi-AZ deployment
http://jenkins.onap.info/job/oom-cd-master/2904/console
Manoj Nair
I can see only the DCAE and closed loop related components in the Edge cloud. Is that a reduced scope ?
There is an indication of DCAE/MEC - Does it mean DCAE + other MEC (ETSI) architecture components (MEAO, MEPM-V) ?
I believe the scope of edge cloud should also include support for slice provisioning based on instruction from central cloud. Also it should preferably based on intents with declarative workflows
Additionally not sure if this diagram is just indicative or planned for Casablanca. As per presentation by 5G Use case subcommittee wiki on slicing in 3GPP - there should be an AMF shared by slices and SMF per slice. With this understanding, would like to know the edge cloud scope with respect to slices - I believe a single edge ONAP manages set of slices in edge infrastructure.
In the diagram I see vOLT and vBBU - Would like to know the role of these two VNFs. Here BBU means C-RAN base band unit or Broadband base unit? If yes I guess DU along with CU fulfils that functionality of C-RAN BBU if I am not wrong. With vOLT I believe this is meant for a backhaul transport link. Is this a correct understanding ? Is this going to be a R-vCPE/E-vCPE kind of connectivity between Edge and Central cloud (management/control plane) ? Is that going to be a separate infrastructure service outside the scope of edge cloud ONAP ?
Raghu Ranganathan
>> scope of edge cloud should also include support for slice provisioning based on instruction from central cloud.
So far, we have focused on IaaS API from onap-central to edge CP. If that IaaS API requests a 'slice', then, edge CP would be expected to support that.
>> if this diagram is just indicative or planned for Casablanca.
not using that diagram. We have gliffy instead.
>> BBU means C-RAN base band unit or Broadband base unit? If yes I guess DU along with CU fulfils that functionality of C-RAN BBU if I am not wrong.
yes.
>> With vOLT I believe this is meant for a backhaul transport link.
no. for fixed access loop. Can be for Residential/Enterprise use cases.
Manoj Nair
Thanks Raghu. The sequence diagram also seems to be bit confusing. From the sequence diagram it gives an impression that Edge ONAP stack is brought up on demand on the Cloud Provider infrastructure. Is this a right understanding ? If this is correct the assumption is that lifecycle and connctivity of EDGE ONAP is managed by Central ONAP.
In this case I did not quite get the sequence of Service Order decomposed into Service Request results in an action on OOM. Are you suggesting changes to SO to have an interaction with OOM ? Or this is going to be a call on Multi-VIM by SO based on the Service Descriptor (for Edge ONAP bring up) ?
If the above understanding is correct, Edge ONAP is similar in nature to a VNF packaged in a Service - Is this a correct understanding ? I am asking this because this will have an impact on how we define the minimum viable product for Edge ONAP - If we bring up Edge ONAP as a VNF on cloud provider infrastructure how can we model the specific components (A&AI, DCAE etc) to be loaded in the Edge ONAP ?
Raghu Ranganathan
>>Edge ONAP stack is brought up on demand on the Cloud Provider infrastructure.
yes...but I would expect this only when the first instance of the edge domain is being used.
>>did not quite get the sequence of Service Order decomposed into Service Request results in an action on OOM
we will need to tweak the sequence, ie., if not already deployed for a service that requires edge-onap components, then request OOM to trigger deployment, etc.
>>Edge ONAP is similar in nature to a VNF packaged in a Service
an edge-onap instance would be expected to support many services deployed in that edge domain. So, the intent was not to do this per service request, etc.
btw, we have not reviewed the sequence in detail during calls. hopefully, we can get to it in a future call.
Bin Yang
The "Sequence Diagram" is a little bit confusing when correlating it to the description "ONAP Edge MVP for Casablanca ". The diagram indicate with Option B there will be close loop on edge infrastructure, while the description tells only Option C support that.
Another question is: How DMaaP deployed with edge ONAP is related/collaborated with the one deployed with centralized ONAP? I mean, is it possible for some ONAP component (e.g. DCAE app ) deployed in centraizlied ONAP to subscribe a topic to DMaaP running within the edge ONAP?
Thanks
Raghu Ranganathan
>>diagram indicate with Option B there will be close loop
..but, not in 'red color font'. See text on right about red text.
for your other Q, see sub-bullets in text for 'key changes needed for option B'. this needs further discussion.
Dominic Lunanuova
May I suggest the following format for section ONAP Edge MVP for Casablanca:
5G - Optimization of the Deployed Network Slices: https://wiki.onap.org/pages/viewpage.action?pageId=28382490
Note:
Note:
ramki krishnan
Certainly Dominic, we are already working on it.
Arash Hekmat
Edge Cloud Team,
I believe there is a significant work item that is missed here which is Edge Modelling in the Central ONAP (for the Modelling Group). From the perspective of the Central ONAP, an end to end 5G service would consist of a set of Edge Clouds connected together through a Central Cloud. For an end to end 5G service to be modeled in the Central ONAP, there needs to be an Edge Descriptor defined in the SDC (just like the VNF/PNF Descriptor today). The Edge Descriptor would define exactly how every part of the Central ONAP would communicate with the Edge entity whatever that entity may be: ONAP Edge, Akraino, ONF CORD, etc. The Edge Descriptor would also define how the Edge would be connected to the Core. Then at design time, for every Edge instance, an Edge Descriptor would be on boarded into the SDC of the Central ONAP to be used in end to end service definitions (templates).
Raghu Ranganathan
this would be part of the 'IaaS' work....in a future release .
Arash Hekmat
But without this how would the end to end service be modeled in the Central ONAP is Casablanca?
Without an Edge Descriptor to abstract out the Edge, every VNF/PNF (DU, CU, ...) at the Edge would have to be modeled in the Central ONAP.
Raghu Ranganathan
we don't....b'coz static set up. .. ....unless, of course, we want to be ambitious and do stuff you suggest. see 'grey region' in sequence diagram.
Srinivasa Addepalli
There is one component called ESR. One way is to enhance ESR (API and A&AI schema representation) to support edge specific configuration.
Raghu Ranganathan
I think you are assuming the problem statement of onap-central managing LCM of each edge node. Do we want to take that on in this work?
Raghu Ranganathan
In other words, we want the edge cloud provider to use onap for their internal use.
Srinivasa Addepalli
Edge LCM (First time, upgrades and patches) involving node/switch software installation/upgrades/patches is not part of ONAP. Normally, it is done using installers (such as compass, APEX, Airship etc..). Once the edge is deployed via installers, it needs to be made aware to ONAP for service orchestration. I was referring to ESR in that context of making edges visible to ONAP. I hope this clarifies .
Raghu Ranganathan
hmm...i would expect the edge node to register with edge provider's systems. what we have assumed so far is the edge provider will expose a service catalog.
Srinivasa Addepalli
Hi Raghu, Looks like I confused you. ONAP will only see the edge-cloud with respect to services, compute profiles (type of edge nodes for HPA). Individual nodes are not visible to ONAP.
In case of Openstack based edge, ONAP will see services exposed by Openstack.
In case of K8S based edges, ONAP will see API server endpoint and OVN North-DB endpoint etc...
We are on the same page ( I hope )
Raghu Ranganathan
ok.
do we want to expand scope of work to include edge provider using onap for their internal domain orch?