Crowdsourcing Available Parking and Available Storage Using Smartphones
Nandugudi Sathyaraja, Anandatirtha
MetadataShow full item record
Smartphones have emerged as the most significant large-scale mobile platform in computing history. However, the scale of smartphone experimentation has lagged behind. Keeping pace requires new facilities that enable experimentation at a scale large enough to ensure that research discoveries translate to the ever growing network of smartphone devices. In this dissertation, we introduce PHONELAB, a smartphone testbed that is open for public experimentation. To demonstrate the efficacy of PHONELAB, we present results from three research studies conducted on PHONELAB with the experiments that ran on participants phones. In the first study, we present results from a usage characterization experiment that ran on 115 phones for 21 days to demonstrate the power of PHONELAB for systems research. In the second study, we present PocketParker, a crowdsourcing system that uses smartphones to predict parking lot availability. PocketParker does not require explicit user input or additional infrastructure and can run effectively without the phone leaving the user's pocket. We consider PocketParker to be an example of a subset of crowdsourcing that we call pocketsourcing . Users interact with PocketParker only when looking for parking spots. PocketParker detects arrival and departure events by leveraging existing activity recognition algorithms. Detected events are used to maintain per-lot availability models allows the PocketParker server to respond to client availability queries. By estimating the number of hidden drivers --those not using PocketParker--we can use a small fraction of monitored drivers to estimate arrival and departure rates and make accurate predictions. Our evaluation uses multiple data sets to determine the accuracy of each PocketParker component and the system as a whole. We show that PocketParker quickly and correctly detects parking events, and that our availability estimator is accurate and robust to the presence of hidden drivers. Finally, we deploy a prototype and use camera monitoring of several parking lots to demonstrate PocketParker's performance in the wild. In the third and the final study, we present the PocketLocker personal cloud storage system. PocketLocker creates scalable, reliable, and performant personal storage clouds out of available space distributed across multiple personal devices. Designed to store rarely-changed files on both interactive devices with limited storage (such as smartphones) and non-interactive devices with large amounts of storage (such as storage appliances), PocketLocker differs from previous systems in not requiring that each device be able to store all available content or be configured to only view certain files. Instead, a storage orchestrator running as a cloud service distributes erasure-coded file chunks across all available devices to attempt to maximize performance and capacity and minimize energy usage at battery-powered clients while meeting configurable backup requirements. And unlike current cloud storage options, PocketLocker is free and will scale as users add devices to their personal cloud. We motivate PocketLocker's design by analyzing two months of file access traces taken from 100 smartphones, and evaluate its performance both using trace-based simulations to explore design parameters and measurements of a prototype Android implementation to establish real-world performance. By locating file content close to where it will be accessed by mobile devices, PocketLocker provides low-latency access to large amounts of content. By exploiting mobility and user charging habits, PocketLocker can meet backup requirements without draining the smartphone's battery.