KubeCon + CloudNativeCon North America 2024

Bloomberg provides an on-premises Data Science Platform (DSP) using cloud-native software to support internal AI model training. It runs on Kubernetes clusters spanning multiple data centers and featuring a diverse range of GPU types. However, managing such a large-scale and heterogeneous GPU environment poses many challenges, such as improving resource utilization, reducing operational costs, and scheduling workloads across different GPU types. In collaboration with the Karmada community, Bloomberg's DSP team has aimed to tackle these challenges by addressing multi-cluster batch job management problems. This talk will delve into the approaches the team has adopted, including: - Intelligently scheduling GPU workloads across multiple clusters - Using Karmada's resource interpreter to support Kubernetes Custom Resource Definitions (CRDs) on top of a multi-cluster architecture - Building a highly available Karmada control plane - Establishing a consistent training job submission interface