Reconciling cloud and mobile computing using activity-based predictive caching

Abstract

Cloud computing has greatly increased the utility of mobiledevices by allowing processing and data to be offloaded, leaving an interfacewith higher utility and lower resource consumption on the device.However, mobility leads to loss of connectivity, making these remoteresources inaccessible, breaking that utility completely during offline periods.We present a concept for reconciling the fragile connectivity ofmobile devices with the distributed nature of cloud computing. We predictperiods without connectivity on the mobile devices before they occurand cache process states for applications running on distributed cloudback-ends. The goal is to maintain partial or full utility during offlineperiods, and thereby to enable an improved user experience. We demonstrateprediction must include real-time behavioral information in additionto location and temporal models. The approach is implemented formobile phones which learn to quantify human behavior using activityrecognition, and then learn patterns in that behavior which lead to disconnectivity.We evaluate it for a streaming music scenario, where datais cached before the user goes offline, allowing seamless playback. Theresults show that theoretically we can successfully predict 100% of disconnectionevents on average 8 minutes in advance (std. dev. 46 secs.)with minimal false-positive caching in this scenario, although in the wildthese events could prove more difficult to predict.