RDMA Programming Design and Case Studies - For Better Performance Distributed

on YouTube ↗06m

" summary

RDMA applied to MXNet over InfiniBand using ibverbs cuts network latency to about 1/5 of IP-over-IB, but requires more code (about 600 lines vs 300 for TCP/IP). Reworking the parameter‑server path reduces memory copies from four to two (4 copies → 2 copies per push/pull), uses RDMA write only, polling/interrupt tradeoffs, ring buffers and a separate completion thread, and yields a measured 1.5× improvement over the first implementation.

AI-written summary · about

" chapters (6)

1. 0:00Introduction
2. 0:24Arguments
3. 1:08Application
4. 1:51Data Row
5. 2:41Implementation Techniques
6. 4:54Conclusion

" related talks

• 18mRDMA is Turing complete, we just did not know it yet!Marco Canini, Simon Peter, Waleed Reda, Dejan Kostić · 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22)
• 23mMulti-Path Transport for RDMA in DatacentersGuo Chen, Peng Cheng, Yongqiang Xiong, Yuanwei Lu, Kun Tan, Thomas Moscibroda, Bojie Li, Jiansong Zhang, Enhong Chen · NSDI '18
• 14mODRP: On-Demand Remote Paging with Programmable RDMARong Chen, Haibo Chen, Xingda Wei, Zixuan Wang, Jinyu Gu, Hongrui Xie · NSDI '25
• 01h 09mAlgorithms for practical distributed agreementNaama Ben-David · Hydra 2021
• 12mPreview: NetworkingAkshay Narayan · OSDI '21 / USENIX ATC '21
• 18mFileMR: Rethinking RDMA Networking for Scalable Persistent MemoryJoseph Izraelevitz, Jian Yang, Steven Swanson · NSDI 2020
• 14mUnderstanding RDMA Microarchitecture Resources for Performance IsolationJitendra Padhye, Xinhao Kong, Jingrong Chen, Wei Bai, Yechen Xu, Mahmoud Elhaddad, Shachar Raindel, Alvin R. Lebeck, Danyang Zhuo · NSDI 2023
• 19mAI/ML Networking Challenges: The Fast and the Finicky!Lerna Ekmekcioglu · DevOpsDays Austin 2025