CloudSLAM: Edge Offloading of Stateful Vehicular Applications
Kwame-Lante Wright, Ashiwan Sivakumar, Peter Steenkiste, Yu Bo, Fan Bai
Abstract
Vehicular applications are becoming increasingly complex and resource hungry (e.g. autonomous driving). Today, they run entirely on the vehicle, which is a costly solution that also imposes undesirable resource constraints. This paper uses Simultaneous Localization and Mapping (SLAM) as an example application to explore how these applications can instead leverage edge clouds, utilizing their inexpensive and elastic resource pool. This is challenging as these applications are often latency-sensitive and mission-critical. They also process high-bandwidth sensor data streams and maintain large, complex data structures. As a result, traditional offloading techniques generate too much traffic, incurring high delay. To overcome these challenges, we designed CloudSLAM. It partitions SLAM between the vehicle and the edge. To manage the complex, replicated SLAM state, we propose a new consistency model, Output-driven Consistency, that allows us to maintain a level of consistency that is sufficient for accurate SLAM output while minimizing network traffic. This paper motivates and describes our offloading design and discusses the results of an extensive performance evaluation of a CloudSLAM prototype based on ORB-SLAM.