Secure Distributed AI Inference Workloads on Slingshot Kubernetes Clusters - Utz-Uwe Haus (HPE)

Описание: Traditional HPC systems offer various levels of job isolation, including secure RDMA enclaves, but often assume a ‘friendly neighbor’ shared batch system environment outside of that. Creating end-to-end attested workflows on them is still a novel development. Kubernetes clusters, on the other hand, offer more isolated execution environments, but network separation typically ends at the VLAN level.
We present a method to execute workloads in a attested environment using RDMA and IP network separation at the linux namespace level on HPE Slingshot. In a setup where K8s is used to run elastic inference workloads.