Share GPU clusters across teams with isolation and fairness using Amazon SageMaker HyperPod
aws ·
This article presents a reference architecture for enabling multiple teams to share Amazon SageMaker HyperPod EKS clusters while maintaining strict isolation and fairness. It utilizes AWS IAM Identity Center for centralized authentication, federating with external providers like Microsoft Entra ID to manage user identities securely. Authorization is handled through per-team SageMaker AI domains and Kubernetes namespaces, ensuring that each team’s workloads are logically separated within the cluster. HyperPod Task Governance plays a critical role by enforcing compute quotas and scheduling priorities to prevent resource monopolization and ensure fair access among competing teams. The architecture also incorporates network policies and POSIX-compliant file systems to further isolate data and traffic between teams. For financial accountability, Kubecost is integrated to provide namespace-level cost allocation, allowing organizations to track and charge back GPU usage accurately. This comprehensive approach allows diverse groups, such as data science and research teams, to collaborate efficiently on shared expensive hardware without compromising security or operational independence.
aws