Live Q&A Session #1 – Srilakshmi Adusumali
summary
Optical link flaps are the common failure mode and teams deploy debouncing, DSP/vendor fixes, proactive monitoring and automatic removal so repeated offenders get taken out; DSF detects and propagates failures in <1 millisecond detection versus front-end ECMP link scans of ~12–25 ms. DSF uses cell-based packets and credit‑based congestion control with tuned credit profiles and PFC thresholds to balance latency and bandwidth; practical limits include 2000-node latency islands and no long-distance RDMA support on OCI.
AI-written summary · about
chapters (13)
- 0:00Intro
- 1:45RDMA
- 3:38Optical link flaps
- 6:00Tradeoffs of ATSW mesh
- 7:30Cellification for DSF
- 9:05Avoiding endtoend PFC
- 10:23Under subscription vs ATSW
- 11:50Performance tuning for DSF
- 14:25Does input balance mode introduce tradeoffs
- 15:25AI use cases willing to trade latency for scale
- 17:10Key bottlenecks when training large models
- 18:30Control plane protocols
- 19:53Credit based congestion control
related talks
32mLive Q+A 2
25mResilience for AI Workloads at Scale: The Fast and the Finicky!
23mLive Q&A with Speakers – James Zeng
27mLIVE Q&A #4
04mLoad balancing through the years
21mQ&A Session 1 – Neil Spring
25mResQ: Enabling SLOs in Network Function Virtualization
15mFiDe: Reliable and Fast Crash Failure Detection to Boost Datacenter Coordination