Kubernetes for Multi-Host Training and Inference - Workload Aware Scheduling

Learn about advanced Kubernetes scheduling strategies specifically designed for AI/ML training and inference workloads in this 38-minute conference talk from CNCF. Discover how current pod-by-pod scheduling frameworks create challenges for AI/ML workloads that require tight coordination between pods, including difficulties with all-or-nothing scheduling and topologically aware compact placement during initial scheduling, failures, and preemptions. Explore innovative approaches being developed by SIG Scheduling to optimize Kubernetes for multi-host AI/ML workloads, understand the specific requirements and challenges these workloads present, and participate in discussions about design solutions that could make Kubernetes the premier platform for running distributed machine learning tasks.