PyTorch: Paco Xu & Kang Zhang on Kubernetes DRA v1.36 for AI/HPC
Paco Xu and Kang Zhang present the Kubernetes Dynamic Resource Allocation (DRA) architecture in v1.36 on the PyTorch official channel, covering GPU, NIC, and NUMA topology-aware scheduling, along with an IMEX daemon case study.
ImportanceLocalEvidenceE2 unreplicatedWrite-upQuick
On September 24, 2026, the PyTorch official channel released a video presented by Paco Xu and Kang Zhang, detailing the architectural evolution of Kubernetes Dynamic Resource Allocation (DRA) in version 1.36. The session addresses the complex requirements of AI and High-Performance Computing (HPC) workloads for GPUs, Network Interface Cards (NICs), NUMA locality, and PCIe/fabric relationships, demonstrating how DRA unifies handling across entry points, resource modeling, scheduler allocation, binding, kubelet preparation, status, health, and access boundaries.
Key topics include topology-aware scheduling for supernode systems, where NUMA, PCIe, NVLink, or fabric placement determines performance. Additionally, the presentation features a case study on IMEX daemon/channel assignment: the design evolved from distributed status updates with conflicts to centralized updates, and finally to distributed updates on a topology-aligned state object, reducing assignment latency from minutes to seconds at thousand-device scale.