A queued change for the Linux kernel's extensible scheduler class sched_ext will prioritize idle SMT siblings when selecting CPUs for task placement. Currently, sched_ext falls back to any CPU within the same last-level cache when no fully idle core is available. The new behavior prefers the idle SMT sibling first, as migration is cheap and preserves L1 cache locality while reducing wakeup latency. NVIDIA's Andrea Righi measured a consistent 2–3% throughput improvement on CPU-bound workloads. The patch is in the sched_ext.git 'for-next' branch and is expected to land in the Linux 7.1 merge window next month.
Sort: