KubeCon + CloudNativeCon Europe 2025

In the era of large models, as models and data are becoming increasingly larger, LLM workloads have extremely high requirements for network throughput and latency.

However, Kubernetes has no awareness of either the parallel models of LLM workloads or the underlying high-speed network communication topology, which leads to a loss in training performance. Meanwhile, many expensive high-performance underlying resources are not utilized more efficiently.

As one of the important projects for Cloud-native AI, Volcano has conducted in-depth research over the past year. It has remodeled the workloads in large model training and inference scenarios as well as the new network topologies, and designed and implemented high-performance scheduling features.

This talk will cover：
1. The complexities related to intelligent scheduling, improving performance and cost-effective
2. Methodology to reconsider the resource model and LLM workload
3. Enhancement to Volcano to optimize training for AI/ML