Qi Alfred Chen University of Michigan QoE Doctor: Diagnosing Mobile App QoE with Automated UI Control and Cross-layer Analysis Qi Alfred Chen University of Michigan
Challenge How can we effectively and systematically study the QoE of popular mobile apps? How can we analyze mobile app QoE which is affected by factors at many layers of the system and the network.
Design QoE Metrics User-perceived latency Mobile Data consumption Energy consumption
Design QOE-AWARE UI CONTROLLER Application control Control the app UI through the InstrumentationTestCase API. (re-signed APK file) Follow a see-interact-wait paradigm. Require some familiarity with Android UI View classes
Design QOE-AWARE UI CONTROLLER App-specific Control Design and User-perceived Latency Collection Facebook Upload post Pull-to-update YouTube Watch video(initial loading time & rebuffering ratio) Web Browsing Load web page(page loading time)
Design QOE-AWARE UI CONTROLLER Data Collection Application Layer Data Collection Collected by the wait component Transport/Network Layer Data Collection Collected by tcpdump RRC/RLC Layer Data Collection Collected by QxDM
Design MULTI-LAYER QOE ANALYZER Application Layer Analyzer User-perceived latency calibration Transport/Network Layer Analyzer Calculate mobile data consumption
Design MULTI-LAYER QOE ANALYZER RRC/RLC Layer Analyzer Obtain RRC state change information from QxDM logs Obtain the power level from Monsoon Power Monitor
Design MULTI-LAYER QOE ANALYZER Cross Application, Transport/Network Layers Identify root causes of QoE problems in the application layer(e.g. device latency or network latency) Cross Transport/Network, RRC/RLC Layers Which RRC state transition cause user-perceived latency Understand how network packets are transmitted in the lower layer
Evaluation Facebook: Post Uploading Time Breakdown Analysis Finding 1: The network delay is not always on the critical path
Evaluation YouTube: Advertisement Impact on Initial Loading Time Finding 9. Advertisements reduce the initial loading time of the actual video, but double the total initial loading time
Evaluation Other findings Finding 2: 3G RLC transmission delay contributes more than expected in the end-to-end photo posting time Finding 3. Facebook’s non-time-sensitive background traffic adds non-negligible overhead to users’ daily mobile data and energy consumption Finding 4. Changing one Facebook configuration can reduce mobile data and energy consumption caused by non-timesensitive background traffic by 20% ……(10 findings in total)
Evaluation Tool Accuracy and Overhead
Conclusion Built a tool, QoE Doctor, which can automatically and repeatly collect QoE data. Analyzed QoE metrics across multi-layer, and diagnosed QoE problems and identify the root causes. Measured QoE metrics in popular Android apps, and quantify how various important factors impact these QoE metrics.