Fewer Cores, More Hertz: Leveraging High-Frequency Cores in the OS Scheduler for

on YouTube ↗22m

" summary

Identifies a concrete problem—frequency inversion—caused by per‑core DVFS and long frequency transition latencies (measured 29 ms rise and 98 ms fall), demonstrates it during an 80‑core Linux kernel build, and implements two scheduler changes in Linux: a cheap local placement and a deferred‑migration policy. Evaluation on 60 applications shows the deferred policy Smove improves performance by >5% for 23 apps (up to 56%) while hurting only 3 (≤8%), and reduces runtime and CPU energy in the case study.

AI-written summary · about

" chapters (14)

1. 0:00Intro
2. 0:40Dynamic Frequency Scaling Before
3. 2:15Dynamic Frequency Scaling Now
4. 3:05Previous Work
5. 4:18Case Study: Compiling Linux
6. 5:20Case Study: Tracing the Frequency
7. 6:25Case Study: Zooming In
8. 7:25Frequency Transition Latency FTL: Latency between a change of load and change of frequency We measure it from idleness to 100% load on our server
9. 9:05Tracing Scheduler Events
10. 9:57Problem: Frequency Inversion Long FTLS
11. 11:06Solution 1: Local Placement
12. 14:07Solution 2: Deferring Thread Migrations
13. 17:11Performance and Energy Evaluation
14. 20:25Take away Frequency inversion problem