Preview: Networking
summary
Explains that RDMA lets a NIC write directly into memory so the CPU does no packet processing, lowering CPU overhead but requiring fancy NICs and a carefully configured network; contrasts RoCE (RDMA over Commodity Ethernet) with iWARP and notes Microsoft’s at-scale deployment (Sitcom 2016) as a working proof. Surveys data center tradeoffs like per-packet hashing versus per-flow paths for load balancing and debates congestion control universality versus domain-specific or ML-driven approaches across five to six orders of magnitude of latency.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:38Goals
- 1:17What is RDMA? Remote, Direct Memory Access
- 2:17RDMA Origin Story
- 3:41Why RDMA?
- 4:44RDMA Choices
- 5:31RDMA Questions
- 6:40Sessions
- 7:22The Dream: One Big Switch
- 9:11Congestion Control
related talks
01h 24mPreview Session
18mRDMA is Turing complete, we just did not know it yet!
18mFileMR: Rethinking RDMA Networking for Scalable Persistent Memory
15mSecuring RDMA for High-Performance Datacenter Storage Systems
02h 53mPreview Sessions
14mODRP: On-Demand Remote Paging with Programmable RDMA
20mEmpowering Azure Storage with RDMA
25mRDMA @Scale