Workshop: Workshop on High Performance Fabrics for AI and HPC
Authors: Sadaf R. Alam (University of Bristol, Bristol Centre for Supercomputing); Jake Watson, Abdelwahab Kawafi, David John, Thomas Green, Matt Williams, and Christopher Woods (University of Bristol); and Ian Johnson, Ellis Pires, Jess Jones, and Utz-Uwe Haus (HPE)
Abstract: AI and HPC increasingly demand interconnects that are simultaneously high-performance, open, and cloud-native. HPE’s Ethernet-based Slingshot fabric substantially shaped the Ultra Ethernet Consortium’s 1.0 transport, establishing it as a proven foundation for open, standards-based RDMA. Exposing its kernel-bypass RDMA plane to multi-tenant, bare-metal Kubernetes is non-trivial: the fabric lives outside conventional Linux networking, and NIC particularities such as resource pools must be accommodated. Prior approaches depend on brittle, loosely-coupled external device plugins, chained CNI plugins, controllers, and databases. Building on a deployment on Isambard-AI, the UK’s sovereign AI Research Resource, we present a Cray eXascale Interconnect (CXI) Dynamic Resource Allocation (DRA) driver making Slingshot devices first-class, scheduler-visible Kubernetes resources. It enables declarative, topology-aware allocation and native multi-tenant isolation without privileged pods, and coordinates cluster-wide Virtual Network Identifier lifecycle through namespace attachment. We evaluate distributed LLM-serving workloads, demonstrating bare-metal RDMA performance with secure isolation within a Trusted Research Environment.
Back to Workshop on High Performance Fabrics for AI and HPC Archive Listing
Back to Full Workshop Archive Listing