You are on page 1of 4

Moving ML Models From Central Into The Edge

NWDAF

Exported on  12/21/2022


NWDAF  –  Moving ML Models From Central Into The Edge

Table of Contents
No headings included in this document

 –  2
NWDAF  –  Moving ML Models From Central Into The Edge

Introduction
In NWDAF we are running prediction models on top of a different use cases which running offline on the central
side.
The training of those models are running and saved offline in a package on the central, per of each use case the
models moved forward into the edge flow to provide the prediction in real-time.
UC Mobility (Gaming) is the first use case.

Link to Epic: 
https://jiradc2.ext.net.nokia.com/projects/NWDAF_NWDAF_ART/issues/NWDAF_NWDAF_ART-76?
filter=myopenissues

Main Architecture Principles:


• Elastically
• All processes are elastic (more workers, more loaders)
• Google storage as a storage of the packages
• Save the history packages
• Interfaces:
• Central - Cloud storage as a source
• Edge - Inference
• Flexibility 
• One generic job with the ability to control the ML-model deployment process (features) inside it by:
• Name of the pipeline + short discerption
• Services \ instances
• Features
• Clone
• Activate \ Deactivate
• Versions
• In every deployment including a short description of the changes
• Ability to roll back
• To the last version easily and fast (need to define "fast" with Yaron) when needed
• To any version in the past
• Performance
• Time running of each pipeline will be no longer then X min\sec\ms
• Depend on: 
• Volume of the package - TBD
• Instances type (type, size, cpu, Io, etc)
• Scheduling
• Ability to chose the time running once and to change it when needed
• Monitoring
• Sending alert by Email\SMS if the ML-model deployment process failed
• Sending alert by Email\SMS if the ML-model deployment process has any delay on time (start\end\on
running)
• Sending logs (need o be align with the different failures, the goal is to find and understand the
problem asap)
• Concurrency 
• Ability to run the ML-model deployment process in parallel 
• Consider instances, data, etc

 –  3
NWDAF  –  Moving ML Models From Central Into The Edge

Next steps: define the ML model (data and broken by...!!!)


Pipeline Data :
• UE tracking - broken by 
• Country
• Edge
• Time
•  TBD (profile of UEs, etc)
• Every model will be related to specific edge

 –  4

You might also like