Improve CPU Utilization by Designing a New Schedule Algorithm
summary
A Linux-kernel offline scheduling class is implemented with two control interfaces (global ratio and per-CPU handoff), lightweight load balancing and coarse binary accounting; tests on Tencent servers raise CPU utilization from about 20% → ~65–70% utilization while keeping online-service latency and failure rates within acceptable bounds (single-module delays drop from ~20 ms spikes to near‑baseline, overall throughput for offline jobs rises ~4–5×). The change lands in the Tencent kernel and minimal CFS code paths.
AI-written summary · about
chapters (10)
- 0:00Intro
- 3:05Latency Sensitivity
- 5:54Classical Issue
- 7:27Solution
- 9:59Load Balancing
- 12:34Scheduling
- 15:12Generic Interface
- 20:39Off Balancing
- 22:58Conclusion
- 25:06Tencent Linux
related talks
32mImprove CPU Utilization by Designing a New Schedule Algorithm
46mOptimizing Scheduler for Linux Gaming
51mTuning RT Kernel to Improve Schedule Performance with Intel Platform
19mDo nothing fast: How to scale idle cpus?
38mFixing Real-Time Scheduler Throttling in the Linux Kernel- 51mTuning RT Kernel to Improve Schedule Performance with Intel Platform
13mHarvesting Spare CPU Resources in Container Systems
52mEffectively Measure and Reduce Kernel Latencies for Real-time Constraints