RankProp: Learning and Propagating Priorities for Cached Apps in Android Memory Management

Modern smartphones keep background apps to reduce latency during frequent app switching. Under memory pressure, Android terminates cached processes and reclaims pages, yet neither path represents when each resident app is likely to return. This paper presents RankProp, which learns an eviction order over the current resident app set and propagates through Android’s memory management. Unlike its predecessor, which scores apps independently, RankProp uses machine learning to train a neural network ranking model that produces an eviction order over all resident apps at each launch event. Android projects the order onto a cached process to reclaim priority and derives a hint for page reclaim. Across 279 user app usage traces, RankProp enhances the aggregate hit ratio from 0.9172 for the calibrated predecessor to 0.9252 across logical app cache capacities from 5 to 15. In an Android emulator with synthetic apps and sustained memory pressure, propagation to both reclaim paths reduced mean launch time by 4.71 percent relative to Android’s stock recency policy. Within this scope, mobile operating systems’ reclaim paths can consume priorities derived from the app eviction ranking without replacing their native decision units.