SMC-R protocol in Storage Software
summary
SMC‑R implements RDMA for TCP sockets on Linux, integrated into SMC tools and the kernel net/smc tree, and is used in IBM Storage Scale on Linux on Z with Rocky adapters and a 100 Gbit switch. Implementation uses AF_SMC sockets or an SMC preload LD_PRELOAD path. With default 128 KB buffers one stream is slower than TCP; increasing to 512 KB yields ~34 Gbit/s and multiple parallel streams reach ~74 Gbit/s. Notes: TCP fallback occurs if peers lack SMC, UDP is unsupported, and kernel patches fix a high‑connection bottleneck not yet merged.
AI-written summary · about
chapters (10)
- 0:00SMC-R in software-defined storage
- 1:07SMCR enables RDMA connections
- 2:34Open-source SMCR implementation
- 3:19Enabling SMC in apps
- 4:04SMCR tools for monitoring
- 6:37Software-defined storage overview
- 8:23SMCR in IBM Storage Scale
- 9:08SMCR performance bottlenecks
- 11:14SMCR hardware benchmarks
- 12:38Summary and Q&A
related talks
21mTCP ≈ RDMA: CPU efficient Remote Storage Access with i10
18mFileMR: Rethinking RDMA Networking for Scalable Persistent Memory
19mSwCC: Software-Programmable and Per-Packet Congestion Control in RDMA Engine
17mSRNIC: A Scalable Architecture for RDMA NICs
27msRDMA -- Efficient NIC-based Authentication and Encryption for Remote Direct Memory
22mResilient AI Supercomputer Networking using MRC and SRv6
12mPreview: Networking
18mRDMA is Turing complete, we just did not know it yet!
