AI for Network Engineers: Understanding Flow, Flowlet, and Packet-Based LB
6 comments
·April 22, 2025immibis
This is not "AI for network engineers" but rather "Network engineering for AI datacenters". I was expecting to read that a small neural network could be used to direct traffic.
lspears
"Though BGP supports the traditional Flow-based Layer 3 Equal Cost Multi-Pathing (ECMP) traffic load balancing method, it is not the best fit for a RoCEv2-based AI backend network. This is because GPU-to-GPU communication creates massive elephant flows, which RDMA-capable NICs transmit at line rate. These flows can easily cause congestion in the backend network."
skeptrune
Far more practical than I expected. I particularly enjoyed the detailed diagrams.
runjake
For the benefit of others' "imposter syndrome", I'll jump in and say it's not just you (the reader, not the parent). It's a really dense article and was really hard for me to follow.
And yes, the diagrams are great. I now must spend the rest of the day resisting the urge to redo all my network diagrams in this style.
Can anyone grok what diagram software the author is likely using?
ttd
I don't know what the author of this article, used but the current iteration of Vexlio [1] is a good fit for this type of diagram. (I am the developer).
[1] - https://vexlio.com
For a latest reference on AI and machine learning for network engineer please check this book by Javier Antich [1].
Please also check the review here [2]. For what it's worth, the book is listed in the "10 Books Every Network Engineer Should Read" [3].
[1] Machine Learning for Network and Cloud Engineers: Get ready for the next Era of Network Automation:
https://www.goodreads.com/book/show/101180344-machine-learni...
[2] MUST READ: Machine Learning for Network and Cloud Engineers:
https://blog.ipspace.net/2023/02/machine-learning-network-cl...
[3] 10 Books Every Network Engineer Should Read:
https://networkphil.com/2024/05/21/10-books-every-network-en...