Minimize reliance on Kubernetes load balancing for internal communication
Minimize reliance on Kubernetes load balancing for internal communication, using direct pod IPs.
Do not rely heavily on Kubernetes load balancing for internal communication.
For workloads like MPI, pods communicate directly using their IP addresses, with service discovery limited to a one-time lookup at job startup. This is suitable when there is little HTTPS traffic and no need for advanced routing patterns.
Pods communicate directly via MPI over SSH using their pod IP addresses, rather than through service endpoints.
Reduce dependency on Kubernetes load balancing for internal communication by having pods communicate directly via IP addresses, especially for MPI workloads.
Simplifies the communication layer and avoids the overhead of Kubernetes services when direct peer-to-peer communication is sufficient and managed by the application.
- Scaling Kubernetes to 7,500 nodes ↗OpenAI · OpenAI News · Jan 25, 2021 · paraphrased
First seen on Subvolts Jan 25, 2021. Updated Sep 11, 2026. Spotted a problem? Tell us.
- Scaling Kubernetes to 7,500 nodesOpenAI · OpenAI News · Jan 25, 2021
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- ToolOrchard: Open framework for scalable agentic AI modelingAn open-source framework for building and studying scalable agentic AI systems. · 10 min
- Prompt patternTemplate for clinical decision support promptsA flexible template for creating clinical decision support prompts for various patient scenarios. · 5 min
- Anti-patternMisconfiguring AI environments for cybersecurity evaluationsDo not tell models they are in a simulation if they are connected to the real internet. · 5 min
- HeuristicImprove LLM idea generation with set-level reinforcement learningTrain models to score idea sets jointly, weighting new ideas by their distance from previous ones. · 120 min
- HeuristicRegenerative software primitives for knowledge preservationUse machine-readable records and regeneration bundles to preserve software knowledge when code is deleted. · 60 min
- HeuristicReasoning-driven synthetic data generation from first principlesConstruct synthetic datasets by reasoning from first principles, not just optimizing individual samples. · 30 min
Questions people ask
- Which AI tools does this work with?
- Any current model. It was seen with in the sources.
- Where did this come from?
- OpenAI (OpenAI News). Subvolts collected, normalized and credited it; the link above opens the original.