Urgent.News

What's breaking now, across hundreds of outlets.

AI

Day 15: AI Network Bottlenecks — Transitioning from iptables to eBPF

Week 3 turns to the network and security layer. Most teams tune the model and the GPU, then discover that the network path inside the cluster is eating CPU that inference needed. This lesson explains why, and what replaces it. How a Service IP reaches a pod When one pod calls a Kubernetes Service, kube-proxy translates the Service IP to a pod IP using Linux iptables rules. Rules are appended to…

On the 15th day of reporting, the focus shifts to the network and security layer of AI workloads. Teams often optimize models and GPUs, only to find that network bottlenecks in the cluster consume valuable CPU resources needed for inference. To address this issue, this installment explains the transition from iptables to eBPF for network path management in Kubernetes Service calls.

When a pod communicates with a Kubernetes Service, kube-proxy translates the Service IP to a pod IP using Linux iptables rules. These rules are appended to chains, and the kernel checks packets against them sequentially, resulting in an O(N) lookup. In a cluster with thousands of Services, this lookup process adds measurable latency to every new connection and consumes host CPU resources.

An eBPF (extended Berkeley Packet Filter) program can map the destination directly in the kernel, bypassing the iptables rule chain. This constant-time lookup operation operates regardless of the number of services, avoiding the performance penalties associated with iptables. eBPF verifies programs in the kernel, ensuring they are safe and reliable to run within the data path without the need for custom kernel modules.

The primary takeaway is that the iptables model exhibits O(N) complexity, making it a CPU tax for AI-intensive workloads. The impact depends on cluster size and traffic patterns, so measuring host CPU metrics and per-connection latency before migrating is crucial. eBPF replaces the rule chain with a constant-time lookup, scaling efficiently with traffic volume rather than the number of services.

While eBPF implementation requires a networking plugin like Cilium, the process is straightforward. To assess the number of iptables rules on a node, use the command `sudo iptables-save | wc -l`. Comparing this count between small and large clusters highlights the exponential growth in iptables rules as service count increases.

The accompanying Python script demonstrates the difference between iptables rule walking and eBPF hash map lookup. By running `python scan.py`, you can observe the difference in lookup times for chains of varying service counts. This example illustrates the performance benefits of adopting eBPF for networking in AI microservices environments.

Rewritten by urgent.news; may contain errors — check the original.

Read the original at dev.to →

Report a problem with this story

More in AI

How much does an AI agent audit trail cost for a small team?

The short answer For a small team, an AI agent audit trail costs $0 to start, $0.99 for one unlimited day, or $29 a month flat, if the trail you need is a record of what the agent did that someone…

  • AI agent audit trails cost $0.99 to $39 per month for small teams
  • Tracing dashboard services include LangSmith and Langfuse, $0.99 to $29/month
  • Evidence trail services like Zambo and ASQAV range $29 to $39/month

Ortho Earth: a WebGPU globe where MapLibre and CesiumJS code run unchanged

I've been building Ortho Earth , a 3D globe map engine that runs in the browser. maplibre × cesium − mercator + α That line is the whole idea: maplibre — code written for MapLibre GL JS runs on it, unchanged, after one import. cesium — code written for CesiumJS runs on it too. − mercator — there is no flat Mercator mode.

  • Ortho Earth runs MapLibre and CesiumJS code unchanged in browser.
  • Engine uses WebGPU, Web Workers, OffscreenCanvas for performance.
  • No API keys, tile servers, or accounts needed for operation.

AI agent tool pricing compared: Lindy vs Gumloop vs Relevance AI vs Bardeen (Nov 2026)

AI agent platforms don't bill like Zapier or Make. Most of them charge in credits that AI tasks burn at different rates, and some add seats on top. I track these tools' official pricing pages every day, so here is what Lindy, Gumloop, Relevance AI and Bardeen actually publish. Last checked Nov 21, 2026.

  • Lindy charges per AI task credits with $50 free daily credits for seven days
  • Gumloop's Pro plan starts at $37/month for 20,000 credits and unlimited agents
  • Bardeen's credits depend on actions: $1 per row created, $3 per enrichment row

More from Saturday 21 November →