Table of Contents
Introduction
Cloud execution of mobile apps—running unmodified apps inside full mobile OSes on commodity servers—is gaining adoption for uses such as GUI agent training and large-scale app testing. These workloads are both temporally dynamic (from second-scale agent interactions to hour-long tests) and highly bursty (up to hundreds of concurrent jobs), demanding an elastic platform that can provision mobile environments quickly and at high concurrency.
Existing stacks fall short because each tenant must cold-initialize the mobile framework (a large collection of shared libraries and system services) before any app can run, inflating boot latency.
We present MCon, the first container system with the framework consolidation architecture: the framework is run as a shared, multi-tenant service instead of a per-tenant component. The core challenge is to practically enable per-tenant isolation on a large framework with no native multi-tenant support. MCon achieves this from outside the framework: by leveraging the fact that apps interact with the framework solely through IPC, we virtualize the IPC interface to construct private views of framework and device resources.
Compared to the best existing stacks, MCon achieves sub-second (15× as fast) cold tenant allocation and improves instance density to 2.6×, decisively improving elasticity for cloud-hosted mobile workloads. MCon has been used to train mobile agents in a major AI company.
Code and Data Release
All the released code and data can be found at our github repo, github.com/mobile-container/mobile-container.github.io.
Figures and scripts
All the evaluation figures used in the paper are included in figs/.
The scripts / jupyter notebooks we used to generate these figures are available in scripts/.
Source code
We have obtained permission to release our full implementation. We are still following relevant internal procedures of the AI company for open source, so the Android source (which is large) may take some time to fully publish.
Other documentation
MCon has been adopted by a major AI company, accelerating end-to-end RL training speed of mobile agents by 56%, and reducing mobile infrastructure cost by 60%. Details of the production system and the deployment experience can be found in docs/production.md.
The table of top-50 apps used in the application benchmark (Section 7.4) is available in docs/fifty-apps.md.
More relevant details (for example, build instructions and guidelines to reproduce our results) are on their way. You can track the current documentation coverage on the docs/README.md page.