Fair Scheduling for AVX2 and AVX-512 Workloads

on YouTube ↗15m

" summary

Modern Intel CPUs drop frequency for AVX2/AVX‑512 code and that slowdown leaks to co‑running tasks. The paper measures a 24.9% average slowdown of non‑AVX work and presents a scheduler change that reduces that gap to 5.4% by disabling vector registers to trap first AVX use, measuring time spent at AVX and non‑AVX frequency bands via counters, and scaling accounted CPU time by the measured frequency ratio. Prototype integrates into Linux scheduling (MuQSS/CFS equivalent) and shows ~4× fairness improvement with negligible overhead.

AI-written summary · about

" chapters (14)

1. 0:00Intro
2. 0:21AVX2/AVX-512 and Fairness
3. 1:31Power-Limited Computing
4. 3:00Example: AVX2 AVX-512
5. 3:55Unfairness in Existing Schedulers
6. 5:53Rethinking Fairness
7. 7:13Frequency Reduction Compensation
8. 7:59Estimating the Frequency Drop
9. 9:38Detecting AVX2 AVX-512
10. 10:59Implementation
11. 12:29Evaluation Setup
12. 13:26Remaining Unfairness (AVX-512)
13. 14:13Overhead
14. 14:27Summary