Building Resilient Network Infrastructure
summary
AI workloads connect 128 to 256+ nodes and demand terabits‑per‑second fabrics with under 10 ms latency; 1 ms extra can add minutes to training and cut efficiency by ~32%. Real failures follow TCP incast patterns that cause packet drops and retransmits; applying RDMA and low‑latency spine‑leaf fabrics prevents collapse. A migration from 10 Gb to 100/400 Gb links plus programmable silicon from Nvidia Broadcom and Intel reduces incidents by 98.5% and scales to petabytes per day.
AI-written summary · about
chapters (7)
- 0:00Introduction and Overview
- 1:14Case Studies and Real-World Applications
- 2:30Modern Network Architectures
- 5:41Network Design Principles
- 7:56Observability and Monitoring
- 8:39Impact of Latency on Performance
- 10:26Conclusion and Key Takeaways
related talks
21mHigh Network Reliability & Availability in FE & BE for Scalable Training Solutions
30mNoisy Neighbors, through Networking
15mHow Our SREs Safeguard Nanosecond Performance—at Scale—in an Environment Built to Fail
35mHigh Throughput with Low Resource Usage: A Logging Journey
26mAndromeda: Performance, Isolation, and Velocity at Scale in Cloud Network Virtualization
34mTroubleshooting Low Latency Application on CNF Deployment
32mActive Fault Finding in Networks
29mMeasuring K8s Network Performance