Breaking the TCP Barrier: Accelerated I/O for S3 With RDMA
Talkat Open Source Summit India 2026 · on Wednesday, 2026-06-17· at 17:15· Lotus Ballroom 2 (Level 3)·
summary
Validation uses SEF + Ceph RGW + RADOS behind NFS Ganesha with Mellanox RDMA NICs (25 Gbps); four clients (each 1 GB) show TCP 426 Mbps vs RDMA 428 Mbps on a 10 GB sequential write and only 2.4% throughput gain at four-way concurrency, but RDMA cuts CPU from 20.6% to 7.8 — a 62% CPU reduction. Reads show ~22% latency/throughput win, small objects (<1 MB) make TCP ~73% faster, and 6.5M ops report zero transport errors.
AI-written summary · about
chapters (13)
- 0:00RDMA and storage acceleration
- 3:35Bypassing TCP bottlenecks
- 7:30NFS Ganesha bridges RDMA and S3
- 9:19RDMA vs local TCP performance
- 10:46CPU reduction with RDMA transfers
- 11:50Concurrent client performance
- 13:15Optimal RDMA scenarios
- 15:41RDMA read performance gains
- 16:45RDMA reliability testing
- 18:10Next RDMA optimizations
- 19:56Breaking the TCP barrier
- 21:10Q&A memory fragmentation
- 22:14Q&A NFS Ganesha bridge
related talks
26mTuning Your Cloud: Improve Global Network Performance for Application
04mCeph and NVMe/TCP
34mEvaluating the Performance and Power Efficiency of Ceph Storage Configurations
39mSuper Fast Real-time Data Processing on Cloud-Native Architecture
26mTuning Your Cloud: Improve Global Network Performance for Applications
40mBenchmarking: Repeatable & Comparable
23m1.2 Tb/S Packet Processing of Radio Telescope Data on a GPU
26mAre You Sure About Your Mesh Performance?