Power Savings & NFV: How to reconcile these 2 constraints ?
summary
Testbed with six cores, Open vSwitch + DPDK, 10Gb ports and T-Rex traffic shows concrete wins: BIOS static low power drops CPU draw to ~9.5 W but cannot reach line rate (14Mpps); default NFV tuning idles ~34–39 W. Core parking and disabling turbo shave watts; fixing frequency at 2.6 GHz hits ~27 W while preserving zero packet loss at test rates. Enabling DPDK interrupt mode yields ~12 W idle and maintains no-loss forwarding; requires NIC/driver support (e.g. x710) and c‑state + cpu‑pinning + libvirt hooks.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:18Challenges
- 1:03cardboard footprints
- 2:09small system
- 2:28gather power consumption
- 3:24static low power
- 4:16static high power
- 5:04live demo
- 11:52whats next
- 13:02software optimization
- 13:25pipeline complexity
- 13:45pipeline depth
- 14:10Conclusion
related talks
26mAre You Sure About Your Mesh Performance?
40mAchieving the Ultimate Performance with KVM
49mPanel Discussion: Moving Towards...
32mImproving High Throughput Computing’s Energy Efficiency - a Measurement-Based Case Study
31mHigh-Performance Cloud Networking With Calico and Vec...
15mOptimizing Application Performance on Kubernetes
31mMeasuring the Impacts of the Preempt-RT Patch
22mThe Internet is Getting Emptier; That's a Sustainability Problem