Live Q+A 2
summary
Operators use bandwidth expectations tied to BDP so large transfers hit about 10% lower than line rate and hop count (leaf vs two‑hop) sets per‑pair targets; redundancy uses dual planes so a server with eight accelerators spans 16 switches / 15 links and can switch planes with zero downtime for single link failures; bit flips get detected by automated loopback tests and on‑box agents that triangulate faults; impact is tracked by an SLI pipeline into a meantime to mitigation metric that drives draining, restarts, and prioritization.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:27Bandwidth and performance
- 1:11Detecting switch failures
- 2:35Tier redundancy in Alibaba
- 4:41Lessons from complex networks
- 7:51Bit flips in networks
- 9:56HPN Sadoo improvements
- 12:45Slow ranks in AI
- 14:53Prioritizing network issues
- 18:29Failure deduction in comms
- 21:43Mitigation impact quantified
- 23:52Network congestion
- 27:26Monitoring effectiveness
- 28:49Network observability without NIC
- 31:00Closing remarks
related talks
23mLive Q&A with Speakers – James Zeng
22mLive Q&A Session #1 – Srilakshmi Adusumali
22mLIVE Q&A #3
15mLive Q&A with Speakers – Joseph Provine
27mLIVE Q&A #4
18mLive Group Q&A Session
17mBackdraft: a Lossless Virtual Switch that Prevents the Slow Receiver Problem
19mAI/ML Networking Challenges: The Fast and the Finicky!