Anuket Project

2021-08-06 AI/ML for NFV Meeting Minutes


Attendees

Sridhar Rao

Steve Casey

Beth Cohen

Kanak Raj

Jahanvi Ojha

Akanksha Singh

Ildiko



Sl. No.TopicPresenterNotes
1Project Status

Finally Approved. 

Thoth will be official name. Gitlab/github/gerrit/jira/slack everything with 'Thoth'

Home Page is update: Thoth

ToDo:

  1. Open LF IT Ticket to get all those things created.

Once it is created, we can start uploading our sources.

2Data Status

Exact File-names for Link and Node Failures

Need updates from Rohit and Girish

3Model Status

Implementation Ongoing.

Target Date: 15 August 2021

Steve Casey:

  1. It makes sense to first start with Decision Trees / Classifiers – Which will make it easy to 'understand and explain' what is happening.
  2. In failure definition, you may have to consider 'planned'/'maintenance' shutdowns.
  3. Failure, especially with VMs, typically occurs due to:
    1. Memory Leaks
    2. Hangs

*** General Tip: For any failure ensure to differentiate true-failure from planned-failure  (ex for link: maintenance/fiber-cut) ***

4

TVLV-Tool

Goal: Create Failure


stress-ng based 

Start-load, Time-Step, Load-Step.

What constitutes the load:

CPU: utilization, operation

Memory: read-size, write-size, read-rate, write-rate

Storage: read-size, write-size, read-rate, write-rate

Network: TCP and UDP.

Beth Cohen: Yes, by doing so, there are 'good chances' it may lead to failure.

Steve Casey:

  1. Another option: run an opensource-VNFs  and, run a Traffic through it. packet-processing: Encrypt/Decrypt

Repeat these for K8S case too.

4

Tools - AlgoSelector

(a) Graph Creation Status

(b) Implementation approach status


  1. Rule-Based Chatbot -  Still not clear-cut choice of Opensource. 
  2. django form - dynamic multi-step form.
  3. CLI library. 

Kanak: 

  1. https://github.com/jdm/asknot
  2. https://spinningup.openai.com/en/latest/

Steve Casey
Reinforcement Learning: Not looked at RL yet. Semi-supervised is used as alternative.

environment parameter: Network State

Adapt: Traffic Policies

Casey: For this usecase, as there is "boundary" for these adaptable parameters, RL is not really reqd., we can use other algorithms/approaches.