At the OSPM 2026 Linux kernel summit, a new sched_ext-based approach for GPU-aware auto-affinitization was presented, achieving up to 80 frames per second on a RegNet workload, outperforming both the default fair scheduler (56fps) and manual numactl pinning (77fps). This work promises to simplify GPU workload management on NUMA systems by automatically optimizing CPU-GPU locality, reducing the need for manual tuning and potentially improving performance for AI and accelerator-heavy applications. The prototype uses a Rust user-space component that queries NVIDIA's NVML library to track per-task GPU utilization, then feeds a BPF map that the sched_ext scheduler (scx_cosmos) uses to migrate tasks to the preferred NUMA node. Caveats include that task memory is not migrated automatically, and aggressive packing may conflict with load balancing.
Background
sched_ext is a Linux kernel feature that allows custom CPU schedulers to be implemented and loaded dynamically using eBPF programs. NUMA (Non-Uniform Memory Access) systems have multiple memory nodes; accessing memory attached to a remote node incurs higher latency. GPU-aware auto-affinitization aims to place tasks close to the GPU they use to minimize data transfer costs.