RDMA @Scale
summary
Azure moves most storage traffic off CPUs: about 65% of data-plane traffic uses RDMA, across links up to 100 km and what the talk claims as an exabyte-scale footprint. The stack uses RoCE v2 over lossless Ethernet plus per‑flow congestion control DCQCN (CNP feedback) to avoid PFC victimization; a validated fluid model tuned router ECN and PFC thresholds, was implemented with Mellanox silicon, and the deployment at 100k‑node scale reportedly works with PFCs rare.
AI-written summary · about
chapters (11)
- 0:00RDMA at Microsoft
- 1:51What RDMA is
- 5:45RDMA in Azure storage
- 8:35InfiniBand to Ethernet
- 11:06Priority Flow Control
- 13:33Victim flow problem
- 16:05Data Center Quantized Congestion Control
- 18:11Tuning DCQCN
- 21:02Mathematical modeling of DCQCN
- 22:25Deployment challenges
- 23:53RDMA future
related talks
20mEmpowering Azure Storage with RDMA
20mRDMA at Cloud Scale: The OCI Experience
15mSecuring RDMA for High-Performance Datacenter Storage Systems
12mPreview: Networking
23mMulti-Path Transport for RDMA in Datacenters
18mRoCE BALBOA: Service-Enhanced RDMA Offload Engine for Data Center SmartNICs
15mFlor: An Open High Performance RDMA Framework Over Heterogeneous RNICs
17mWhite-Boxing RDMA with Packet-Granular Software Control
