Ashish Bijlani

Ph.D. Student

School of Computer Science

Georgia Institute of Technology

Date: Tuesday, March 10th, 2020

Time: 9:00 - 11:00am EDT

Location: Klaus 1212

Committee

Dr. Umakishore Ramachandran (Advisor, School of Computer Science, Georgia Institute of Technology)

Dr. Mostafa Ammar (School of Computer Science, Georgia Institute of Technology)

Dr. Ada Gavrilovska (School of Computer Science, Georgia Institute of Technology)

Dr. Vivek Sarkar (School of Computer Science, Georgia Institute of Technology)

Dr. Raghupathy Sivakumar (School of Electrical and Computer Engineering, Georgia Institute of Technology)

Abstract

Smart mobile devices have largely evolved as primary tools for everyday personal computing needs. This has led to a surge of apps for communication, travel and planning, gaming, health tracking, social networking, home automation, media, and entertainment. Personal mobile devices are also increasingly being used for work. The versatility of these devices, however, poses high, and often conflicting, storage demands. As a result, devices quickly run out of space given their limited storage capacity.  

This work aims to address the storage management challenges on smart mobile devices. We begin by carrying out a large-scale longitudinal study of millions of mobile apps to report increase in their sizes over time.  While historically small in size (25MB), our results show that modern mobile apps have evolved as feature-rich, large packages, each capable of consuming up to 4GB of storage space.  Developers create “build once, run everywhere” apps to target multiple CPU architectures and demographics regions that consume redundant storage, despite Google’s effort to create small apps. We further show that as apps become popular, developers pack more features to create to engage users, resulting in bigger apps. Yet, apps are not constrained by storage quota limits, and developers freely exploit persistent storage by hoarding data and frequently caching content as needed to offer high performance and rich user experience. 

This work further leverages results from a user study in the wild to show that users only use a small fraction of apps and features, and the usage is highly correlated to user context (e.g., location, day, time). Drawing upon our findings, we proposes a context-sensitive quota model for efficient and application-transparent storage management of smart devices. Storage space consumed by contextually unwanted apps/data is temporarily reclaimed by employing multiple techniques, such as selective compression, deletion, content adaptation, offline deduplication, and Cloud-backed hierarchical management. Reclaimed data is reconstructed proactively under predictive usage or served on-demand either by computation on the device or fetching it from the Cloud. To understand the design implications, we have implemented a platform storage service on Android that performs proactive management of local storage. Attributed to our highly optimized framework design, our prototype implementation imposes negligible runtime overhead and can free up significant storage space.