AI
DRANet: RDMA and Topology-Aware Scheduling for AI Nodes
DRANet is a Kubernetes network driver written by Antonio that uses dynamic resource allocation to deliver high-performance networking without a stack of separate plugins. It handles resource discovery, exposing every interface with attributes like RDMA capability and PCI address. For AI training workloads, placing a GPU and NIC on the same PCI route matters for throughput. DRANet enforces that constraint at scheduling time using standardized attributes. Cross-node topology awareness is still in progress, coordinated with SIG Scheduling.
Gateway API Inference Extension: Smarter Routing for LLM Traffic
The Gateway API Inference Extension, also called Inference Gateway, treats LLM serving as distinct from web serving. Inference workloads carry an order of magnitude fewer requests, and every request requires inspecting the body, not the headers. The extension introduces inference-aware scheduling and prefix-cache-aware routing. A new flow-control feature buffers requests against the inference pool to prevent overload and expose pool-fullness metrics for autoscaling. Priority mixing, filling idle capacity with low-priority batch requests that can be evicted, was figured out the day before this talk.
Kubernetes Agentic Networking: Access Policy and Observability for Agents
The Kubernetes Agentic Networking project targets communication between agents, between agents and tools, and between agents and LLMs, covering both ingress and east-west traffic across cluster boundaries. Two alpha CRDs are under active development: XBackend, which defines external inference service endpoints, and XAccessPolicy, which specifies which agents may call which tools. A reference implementation using Envoy extracts the tool name from the JSON-RPC body, places it in dynamic metadata, and feeds it to an authorization policy. A demo talk was scheduled for the following day at 17:30.
AI Gateway Working Group: Payload Processing and Egress Standards
The AI Gateway Working Group, sponsored by SIG Network and SIG Multicluster, focuses on standardizing how gateways handle AI traffic. It builds on existing Gateway API resources rather than replacing them. Two active proposals cover payload processing, which enables prompt guardrails, semantic routing, and RAG augmentation by inspecting the request body, and egress gateways, which standardize routing to external inference services such as cloud provider models. New CRDs are introduced only when existing HTTP routes and gateway resources cannot cover the use case.
Q&A
Does DRANet replace network plumbing or CNI v2? DRANet works in parallel with the network plumbing group rather than replacing it, and the two communities are now collaborating more closely. ▶ 23:53
What are the networking challenges for prefill-decode disaggregation? The current instantaneous decode scheduling used by XRO does not support deferred decode needed for vLLM’s HBM management, so additional components may be required. ▶ 25:22
Does AI networking require changes below the application layer? Running large numbers of agents in sandboxes creates a dynamic environment that may drive new requirements for policy and lower-layer networking. ▶ 29:02
Notable Quotes
inference workloads should have an order of magnitude less request volume, even more most likely. Kellen Swain · ▶ 17:28
All they want is just give me my interface with it as little virtualization in between as possible. Nadia Pinaeva · ▶ 18:20
okay this agent last night had made like 1,000 tool calls to this tool But why did that happen? You need to get the whole picture of that. David Martin · ▶ 22:50
you should use it just don’t use it in production. Bowei Du · ▶ 16:30
Key Takeaways
- DRANet schedules GPUs and NICs on the same PCI route, improving AI training throughput without extra plugins.
- Inference Gateway’s new flow-control buffer prevents pool overload and exposes fullness metrics for autoscaling.
- XAccessPolicy CRD controls which agents call which tools, with Envoy parsing the JSON-RPC body to extract tool names.
About the Speaker(s)
Bowei Du has been a SIG-NETWORK member since 2015 and serves as SIG Network chair. He led scalability work including EndpointSlice and nodelocaldns, co-authored the Kubernetes DNS spec, helped move Ingress to GA, and initiated the Gateway API project.
David Martin has spent over 15 years in managed services, from Node.js mobile backend platforms to API and integration products on OpenShift. At Red Hat he works on MCP gateway.
Nadia Pinaeva is a Senior Software Engineer at NVIDIA focused on Kubernetes networking. She works with the SIG-network-policy group on network security and develops the ovn-kubernetes network plugin.
Kellen Swain is a software engineer at Google and a maintainer of the Kubernetes Inference Gateway project, working on the GKE inference team.
Haiyan Meng is a Senior Software Engineer at Google on the GKE team since 2019. She works on the Kubernetes Agentic Networking project and has broad experience with Config Sync and Fleet multi-cluster solutions.