SYSTEM Cited by 1 source

sched_ext¶

Definition¶

sched_ext is an upstream, BPF-based extensible scheduling framework for the Linux kernel, officially merged in kernel v6.12. It allows scheduler developers to implement custom scheduling policies as BPF programs that the kernel invokes through event-driven callbacks (thread wake-up, enqueue, dispatch, idle transitions). The policy runs as a user-space binary that loads the BPF program into the kernel.

Architecture¶

sched_ext exposes a set of scheduling callbacks to BPF programs:

Thread wake-up: choose a CPU when a thread becomes runnable
Enqueue: place a thread in a run queue
Dispatch: select the next thread when a CPU becomes idle
Idle transitions: respond to CPUs entering/leaving idle states

The policy is packaged as a user-space binary that loads the BPF scheduler program. To update the policy, restart the scheduler process — no kernel rebuild or reboot required.

Origin¶

Co-developed by Meta and Google (authors of ghOSt) to create an extensible scheduling framework suitable for upstream Linux. ghOSt (SOSP 2021) demonstrated the value of user-space scheduling policies; sched_ext is the production-ready, upstreamed realisation.

Production deployment at Meta¶

Meta deployed sched_ext on its ads serving fleet to address a latency regression caused by the CFS→EEVDF transition (Linux 6.4→6.9). The custom policy soft-partitions CPUs into latency-critical and background pools, improving L3 cache locality. Results:

28% p99 latency reduction (initial launch)
+60% additional p99 reduction (two follow-on user-space policy updates)
3.28 MW power savings
+1.1% weighted-ads-ranked
18% fewer timeout errors

(Source: sources/2026-07-13-meta-modernizing-ads-service-open-source-kernel-scheduler)

Significance¶

sched_ext transforms CPU scheduling from a monolithic kernel design decision into a per-workload concern iterable in days. The userspace deployment model makes scheduling optimisation available to any operator — hyperscaler, cloud provider, embedded systems team — without forking the kernel.