Jun Xiao, Qinhui Gu, Ligeng Chen, Lizhi Sun, Zicheng Wang, Yinggang Guo, and Lu Liu, Honor Device Co., Ltd.; Hao Wu, Nanjing University; Borui Li, Southeast University
CPU scheduling performance on mobile devices, especially under user interactions, is hindered by a fundamental semantic gap: the kernel scheduler lacks visibility into the user interaction context, treating latency-critical UI threads and background tasks equally. In this paper, we identify an impossible trinity, i.e., scarce prime cores, cross-process IPC dependencies, and tight latency deadlines, which exacerbate the mobile scheduling problem. To survive the impossible trinity, we present MUSched, a semantic-aware scheduling framework for modern mobile devices that makes interaction capability a first-class scheduling objective. MUSched disentangles cross-process dependencies for critical threads along the interaction path and places these threads in a new VIP scheduling class between RT and CFS, allowing interaction-critical tasks to preempt normal background work without compromising system stability. Furthermore, MUSched proposes a scheduling policy plug-and-play mechanism that facilitates on-demand policy update in the user space without kernel recompilation for COTS mobile devices. In laboratory evaluations, MUSched reduces average application cold-start time by 14.8%. Furthermore, MUSched has been deployed on more than 20 million mobile devices since 2024. The deployment results show that MUSched reduces real-world startup anomalies by more than 30.7%. This deployment underscores the pivotal role of semantic-aware scheduling in achieving optimal mobile Quality of Experience.
