Silicon and Networking
summary
AWS uses custom Nitro silicon for both massive active monitoring and for a new transport that reduces tail latency. Nitro cards run 1.5 billion probes per minute to detect forwarding, latency deviations, and loss and drive automated remediation. AWS invents SRD (scalable reliable datagram) and exposes it via ENA Express transparent upgrade on EC2 so TCP/UDP apps gain multi‑path packet spraying, faster retransmits and reordering. Result: single‑flow throughput jumps (example 5x from 5 Gbps to 25 Gbps), P99/P99.9 tail latency drops up to ~50–83%, 4K live video at 12 Gbps meets <14 ms, Redis tests show ~40% P99 improvement.
AI-written summary · about
chapters (7)
- 0:00Start
- 1:30Nitro Silicon
- 4:28EC2 instance networking timeline
- 5:46Amdahl's Balanced System Rule of Thumb
- 9:58Active Monitoring with Nitro
- 12:02The Customer Challenge
- 14:35Introducing Scalable Reliable Datagram
related talks
51mAdvanced approaches to traffic inspection & network diagnosis w/
25mScaling Up, One Network Bottleneck at a Time
26mTuning Your Cloud: Improve Global Network Performance for Applications
26mTuning Your Cloud: Improve Global Network Performance for Application
35mLive Technology Panel
18mThe Evolution of High Performance Networks: A Storage Perspective and Practices
07mWelcome Remarks
37mPlease Give Me Back My Network Cables! On Networking Limits in AWS