Understanding on-device bufferbloat for cellular upload

Despite the extensive characterization of the growth of cellular network traffic, we observe two important trends not yet thoroughly investigated. First, fueled by the LTE technology and applications involving wearable devices and device-Todevice (D2D) communication, device upload traffic is increasingly popular. Second, the multi-Tasking and multiwindow features of modern mobile devices allow many concurrent TCP connections, resulting in potentially complex interactions. Motivated by these new observations, we conduct to our knowledge the first comprehensive characterization of cellular upload traffic and investigate its interaction with other concurrent traffic. In particular, we reveal rather poor performance associated with applications running concurrently with cellular upload traffic, due to excessive ondevice buffering (i.e., on-device bufferbloat). This leads to significant performance degradation on real mobile applications, e.g., 66% of download throughput degradation and more than doubling of page load times. We further systematically study a wide range of solutions for mitigating ondevice bufferbloat, and provide concrete recommendations by proposing a system called QCUT to control the firmware buffer occupancy from the OS kernel.