Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

Sukhdev Kapur

Vikram Siwach



Project Technical Lead:               Vikram Siwach


Project Committers detail:

Committer

Committer

Company

Committer

Contact Info

Committer Bio Committer Picture Self Nominate for PTL (Y/N)






Sukhdev Kapur

Juniper NetworksSukhdev Kapur


Vikram SiwachMobiledgeXVikram Siwach















Presentation:


MEX-JNPR_BP2.pdf



AttributesDescription

Informational

Type

New Blueprint for enabling AI/ML and low latency AR/VR capabilities at the Edge

New submission for integrating Tungsten Fabric for Edge Cloud for AI/VR applications for developers in a distributed environment
Industry SectorEdge Cloud, Enterprise, 5G, and IoTThe blueprint empowers Integrated Edge Cloud architecture to use single SDN Controller (Tungsten Fabric) for all type of Containers, Virtual Machines, and Bare Metal Servers.
Blueprint Family - Proposed   NameIntegrated Edge Cloud
Use Case

Programmability on switches and I/O Accelerations on programmable NICs & embedded FPGAs to deliver AI/ML workload placement and low latency demands of AR/VR applications onboarding edge stack


Blueprint proposed Name

AI/ML and AR/VR applications at Edge


Initial POD Cost (capex)

Leverage white boxes, standard NICs:  The cost of POD will depend upon the hardware profiles and peripherals desired


Scale & TypeThe deployment will employ baremetal as well virtual machines and containers. The smallest footprint for minimal Cloudlet is 9VMs to a large Cloudlet could span to 8-Cluster Cloudlets
Applications

AI/ML streaming workloads and AR/VR applications


Power Restrictions

Less than 10Kw

Infrastructure orchestration

OpenStack Queens or above

Docker 1.13.1 or above

Container Orchestration –K8s 1.10.2 or above

OS - Ubuntu 16.x, CentOS


SDN

Tungsten Fabric Kernel vRouter, DPDK vRouter, SR-IOV and SmartNIC


Workload Type

Containers over VM or baremetal


Additional Details

● Runs on Commodity HW
● x86, ARM, SoC, ...
● Multiple options for partial or full NIC offloads (Intel, Netronome, Mellanox)
● Future Supports:
● eBPF/XDP offload