Anuket Project
2021-10-29 AI/ML for NFV Meeting Minutes
- Sridhar Rao
- Beth Cohen
Attendees
| Sl. No. | Topic | Presenter | Notes |
|---|---|---|---|
| 1 | EUAG Paper: Action Item Discussions
| Platform: Data Pipeline? ML-Models ? Language/libraries ? Integration-Flexibility ? Project: Explore O-RAN. Data Modeling: (a) Sources (b) time-series (c) logs (d) understanding of the columns (e) Terminologies used (f) naming consistency. Can YANG help ? (To Explore). Does YANG already cover metrics/logs - infrastructure monitoring ? Is there any platform which is normalizing multiple-sources and having a common data model? Repository of the problems (use-cases). Example of where bias in data can have unintended consequences: https://courses.cs.duke.edu//spring20/compsci342/netid/readings/facialrecnytimes.pdf Test/Certify: ML-Model. Against Common Training + Test Data. Metrics: Speed + Accuracy. Publish: Training and expected metrics. Richness of the data-set is very important. Training: Amount of training available on AI/ML and Networking in public is HUGE. LF-Course?\ | |
| 2 | Kubernetes Failure Emulation | Hypothesis: We cannot emulate failure of a pod, by running stress tools. CPU loads have no impact. Even if you allocate 2GB RAM (with static configuration), and do memory operation (stressng) with buffer size of 4 to 6 GB, still there will be no failures. Tried these configurations: https://github.com/opensource-tnbt/stressng-images. What is more important ML-problem for K8S for NFV usecases is an open question ? | |
| 3 | GANs for Synthetic Data Generation |
Ildiko: Open Telemetry meeting (OIF), focus is on kubernetes. Area of Tracing. eBPF/Jaeger. | |
| 4 | BERT for Openstack Log Analysis | ||
| 5 | Webinar, Testing Forum | Role of the Projects such as Thoth.
|