Workshops Archive

Sovereignty above the scheduler: pluggable site policies and a federated view for AI/HPC with Waldur


Workshop: Sovereign AI Supercomputing Cloud Workshop

Authors: Ilja Livenson (OpenNode)

Abstract: Sovereign AI and HPC infrastructure requires that sites can enforce their own policies and replace underlying components without breaking their participation in federations.

We present the architecture of Waldur, an MIT-licensed allocation and accounting platform that serves as the allocation component of the EuroHPC Federation Platform and runs in production at sites including CSCS, Isambard-AI, Cambridge Dawn and the Estonian Government Cloud. Waldur separates two concerns that are often entangled. At the site level, a pluggable agent model lets each site map federated allocations onto its own backends (Slurm, Kubernetes, OpenStack or others) and apply local policies for account provisioning, quota enforcement, usage reporting and access rules. Scheduler choice and local governance thus remain site decisions. At the federation level, a common data model for projects, allocations, users and usage gives funders, allocation committees and users a single view across heterogeneous sites, without requiring those sites to adopt the same stack.

We discuss the design of the plugin boundary: what must be standardised for federation to work, what should stay site-specific, and how this boundary limits the impact when any one component changes its license or stewardship. We close with open questions on shared allocation and accounting standards across sovereign federations.


Back to Sovereign AI Supercomputing Cloud Workshop Archive Listing

Back to Full Workshop Archive Listing