Hi everyone,
Some of you may remember the "Kubernetes Network Drivers" concept we initially presented at the Contributor Summit in Paris last year. I have been working since then on the execution of this idea, collaborating with the Device Management WG in DRA and the OCI and NRI communities to get the necessary changes into the runtime specifications, and developing a reference implementation, DraNet.
I'm really happy to share some results of this work in the following research paper: "The Kubernetes Network Driver Model: A Composable Architecture for High-Performance Networking."
The paper shows how this new architecture solves complex networking challenges,focused on the AI/ML use cases, using an approach that is much simpler than what exists today. A huge advantage of this model is how it reduces the operational complexity, making it far easier to deploy, manage, and debug.
2. Is DraNet a Multus replacement?
No. This is the core philosophy of Kubernetes Network Drivers. DraNet is not a general-purpose tool; it is a specialized driver focused on solving the complex problem of attaching network interfaces (specially RDMA devices) for AI/ML pods. The idea is to enable an ecosystem of composable, function-specific drivers: a MultiNetwork KND, an SRv6 KND, an MPLS KND, .... Composition happens at the functional level, rather than by chaining technologies together.
I look forward to your feedback and discussion.
Best regards,
Antonio Ojea