Honor MUSCHED: eBPF VIP Scheduling Cuts Mobile Touch Latency 31% and Frame Drops 92%

Modern mobile devices face an "impossible trinity": ultra-tight frame deadlines (8.3 ms at 120 Hz), bursty interactive workloads, and traditional Linux schedulers that are either too aggressive (RT) or too generic (CFS/EEVDF). Honor's MUSCHED, presented at Chinasys 2026, addresses this with a semantic-aware scheduler built on sched_ext eBPF.

Core Innovations

VIP scheduling class : A new priority tier between RT and CFS, using FIFO ordering with a 3 ms time slice and per-task quota limits (WebView 120 ms, audio 20 ms, video 10 ms, display 20 ms). Tasks exceeding both slice and quota are temporarily demoted.

Scenario-aware annotation : Function hooks in Android framework (SystemUI, launcher animations, focus changes, frame rendering) dynamically tag critical threads (Animator, UI, Render, WebView Loading) as VIP when they move to foreground.

Cross-process priority propagation : For Binder IPC and lock dependencies, VIP tags propagate to remote service threads (e.g., SurfaceFlinger) and lock holders, eliminating long-tail latency. Tags are cleared after IPC completion or lock release.

Scheduling Mechanisms

CPU Selection

On wake-up, VIP tasks prefer idle performance cores; if none, the least-loaded performance core without RT/VIP tasks; finally, a core with no RT tasks and fewest VIP tasks.

Load Balancing

Pull : Idle CPUs steal the first runnable VIP task from queues containing both RT and VIP tasks, or from the queue with most VIP tasks.

Push : On each tick, if a VIP task has waited >4 ms while an RT task runs, migrate it to a core with no RT tasks and fewest VIP tasks.

Lock Wait Optimization

VIP preemption : VIP threads break FIFO wake-up order, jumping ahead of non-critical waiters.

VIP inheritance : When a VIP thread blocks on a lock held by a normal thread, the holder temporarily inherits the VIP tag to accelerate unlock; tag is cleared on wake-up.

Implementation via sched_ext

sched_select_cpu

: Chooses target core for VIP tasks. sched_enqueue / sched_dispatch: Places tasks into per-CPU VIP DSQs via scx_bpf_dispatch to PRIO-DSQ.

Load balancing uses scx_bpf_consume(DSQ_ID_VIP) for local consumption and stealing from neighbor VIP DSQs.

State tracked via two BPF maps: cpu_contexts ( BPF_MAP_TYPE_PERCPU_ARRAY): per-CPU idle time, load, queued task count, suspend state. task_contexts ( BPF_MAP_TYPE_TASK_STORAGE): per-task weight, execution state, compute load, enqueue/dequeue timestamps.

Evaluation Results

On Snapdragon 8 Gen 4 + Android 15 flagship hardware:

Touch-to-display latency reduced up to 31% .

92% of frame drops eliminated at 120 Hz.

Cold-start case study on Honor Magic 7 (10 popular apps): 300+ threads created, hundreds of MB to GB memory allocated, intense CPU contention. MUSCHED reduced cold-start time, D-state (uninterruptible sleep) and Runnable-state durations significantly (see charts).

Production deployment across >20 million devices shows measurable reductions in abnormal latency for animation, swipe, and startup scenarios.