Open Source Summit North America 2025

This talk presents a method to enhance CPU scheduling by leveraging machine learning (ML) to extract the key features necessary for task migration, allowing for dynamic and stable optimization of workloads in scenarios prone to CPU imbalance. The approach is built on the sched_ext framework, which integrates eBPF to support user-defined scheduling policies within the Linux kernel.

While conventional approaches maximize CPU utilization, they often overlook contention for lower-level hardware resources, leading to performance bottlenecks -- particularly in compute-intensive servers. By using an ML-based, resource-aware load balancer, this method effectively addresses such imbalances. With sched_ext, we can collect training data and run inference on the model without modifying the kernel.

Our experiments demonstrate that, for certain workloads, this ML-driven approach can outperform EEVDF, offering notable performance gains for the CPU scheduler.