TPP: Accelerate Application Launch via Two-Phase Prefetching on Smartphone

Abstract

The fast app launch is crucial to users' experience and it is one of the eternal pursuits of manufacturers. Page fault is a critical factor leading to long app launch latency. Prefetching is the current method of reducing page faults during app launch. Before app launch, prefetching all demanded pages of the target app can speed up the app launch effectively, but it always uses the memory of several hundred MB, leading to low memory and slowing other apps' launch. Prefetching during application launch uses memory effectively, however, current methods are not aware of the order of pages accessed, causing noticeable accessing-prefetching order inversions, which results in limited acceleration of app launch. In order to accelerate the application launch effectively with little memory usage, we propose a Two-Phase Prefetching schema (TPP), which performs prefetching via two phases: 1) Before the app launch, to increase the efficiency of memory usage in prefetching, TPP prefetches few critical pages with app prediction, which is based on Long Short-Term Memory (LSTM) with high accuracy. 2) During app launch, TPP prefetches the rest of the critical pages via an order-aware sliding window method, resolving the accessing-prefetching order inversions and significantly reducing the app launch latency. We evaluate TPP on Google Pixel 3, compared to the state-of-the-art method, TPP reduces the application launch time by up to 52.5%, and 37% on average, and the data prefetched before the target application started is only 1.31 MB on average.