Taming the Chiplet: High Performance CCX scheduling via BPF
summary
Implements fine-grained CCX-aware scheduling using BPF plus a userspace agent that maintains per-CCX runqueue state and runs an asynchronous bin-packing allocator to place VM thread groups. Targets AMD Zen5-style topologies (256 cores across two NUMA nodes, SMT2) with up to 2x overcommit. Heuristics trigger controlled spillover during spikes to minimize spread; offline tuning of bin-packing coefficients and overload thresholds yields double-digit throughput uplifts and much higher CCX locality while moving more logic into BPF cuts amortized overhead.
AI-written summary · about