When Cloud Storage Meets RDMA
summary
Pangu integrates RDMA into production cloud storage and reports concrete wins: a user‑space stack (USS OS and USS file system) boosts CPU efficiency by over 5x; UMR and CSO cut memory bandwidth by >30% and raise single‑node throughput. QP explosion drops from ~22,000 to ~1,500 via a thread‑sharing design at a 0.3 µs cost. RDMA/TCP switching and fast detection (~10 s) contain PFC storms. RDMA‑enabled Pangu runs in production for >3 years, serving shopping‑festival traffic.
AI-written summary · about
chapters (5)
related talks
15mMore Than Capacity: Performance-oriented Evolution of Pangu in Alibaba
18mThe Evolution of High Performance Networks: A Storage Perspective and Practices
25mDisaggregating Persistent Memory and Controlling Them Remotely
14mPD3: Prefetching Data with DPUs for Disaggregated Memory
20mEmpowering Azure Storage with RDMA
25mRDMA @Scale
15mSecuring RDMA for High-Performance Datacenter Storage Systems
28mPersistent Memory plus RDMA, new age remote device
