Skip to content

Spiderpool: Dynamic Topology-Aware RDMA Allocation For GPU-Based AI Workloads#

KubeCon + CloudNativeCon North America 2025 digest Program Core pick 10–13 November 2025

Recording

Weizhou Lan, DaoCloud — Core (Project Lightning Talk)

Play the recording: Spiderpool: Dynamic Topology-Aware RDMA Allocation For GPU-Based AI Workloads

  • Why: A precise physical-topology networking problem — aligning RDMA NICs with PCIe-affined GPUs — that the standard device-plugin/CNI split can't solve. This sits right at the DC-hardware/cloud-native seam the PM cares about.
  • What: A mechanism in Spiderpool to allocate only the PCIe-affined RDMA NICs a pod actually needs, instead of attaching every interface and RDMA device. (from abstract only)
  • How: Built on DRA and NRI, Spiderpool detects the GPU devices assigned at pod-creation time and allocates the matching PCIe-affined RDMA NICs and Ethernet interfaces for GPUDirect RDMA, minimizing overhead while preserving RDMA performance. (from abstract only)
  • Where applicable: GPU nodes doing distributed inference/training where a container needs a subset of GPUs; requires RDMA NICs with discoverable PCIe affinity and DRA/NRI support. Lightning-talk depth.