Alibaba Stellar: A New Generation RDMA Network for Cloud AI#
SIGCOMM 2025 digest Program Core pick Coimbra · 8–11 September 2025
Jie Lu et al. (Alibaba Cloud)
Why: RDMA virtualization (SR-IOV) pain — slow container spin-up, resource limits, traffic steering — is exactly the multi-tenant cloud-AI networking the PM works on.
What: A production RDMA virtualization stack that spins up virtual devices in seconds, cuts container init time 73%, and improves LLM training speed up to 14%.
How: Para-Virtualized DMA (PVDMA) for on-demand memory pinning, an extended Memory Translation Table (eMTT) for GPUDirect RDMA performance, and RDMA Packet Spray for multipath.
Where applicable: Alibaba-specific production setup; the PVDMA/eMTT ideas generalize but the numbers are from their cluster.