Cloud-Assisted Streaming for Low-Latency Applications Xiaoqing Zhu, Jiang Zhu, Rong Pan, Mythili Suryanarayana Prabhu, and Flavio Bonomi Advanced Architecture & Research, Cisco Systems Inc., San Jose, CA 95134, U.S.A. {xiaoqzhu,jiangzhu,ropan,mysuryan,flavio}@cisco.com Invited Paper Abstract—Media cloud services offer an unique opportunity for allevi- ating many of the technical challenges faced by mobile media streaming, especially for applications with stringent latency requirements. In this paper, we propose a novel cloud-assisted architecture for supporting low- latency mobile media streaming applications such as onling gaming and video conferencing. A media proxy at the cloud is envisioned to calculate the optimal media adaptation decisions on behalf of the mobile sender, based on past observations of packet delivery delays of each stream. The proxy-based intelligent frame skipping problem is formulated within the Markov Decisio Process (MDP) framework, which captures both the time- varying nature of video contents as well as bursty fluctuations in wireless channel conditions. The optimal frame skipping policy is calculated using the stochastic dynamic programing (SDP) approach, and is shown to consistently outperform greedy heuristic schemes. Our simulation studies further characterize how system performance is influenced by various key factors, such as application playout latency, network round-trip-time, and wireless link throughput. Index Terms—cloud computing, low latency media streaming, mobile video conferencing, video adaptation I. I NTRODUCTION Recent years have seen a proliferation of smart mobile devices, which, in turn, has fueld the rapid growth of mobile media traffic. Many of the applications supported by today’s mobile devices have stringent latency requirements. Examples include media sharing of live events, online gaming, mobile video conferencing, and media- rich virtual desktops. According to [1], the latency threshold for first-person-based avatar games such as racing and combating is in the range of 50 - 100ms. Whereas for video conferencing, the recommended one-way latency is below 150ms [2]. Packet delivery latency, in addition to bandwidth, has is a key performance metric for such applications. On the other hand, mobile media streaming remains a daunting challenge due to inherently time-varying wireless communication channel, unpredictable user demand in the media cloud, and fluc- tuating source rate of media contents generated on-the-fly. Moreover, the mobile sender typically has limited battery power and compu- tational resources, hence may not afford to implement sophisticated adaptation algorithms for matching the rate of streaming media to available wireless network throughput. Mobile media streaming could benefit from some form of as- sistance from media cloud servers or proxies in many ways. For instance, the relatively abundant and low-cost computational power of cloud servers could be leveraged for carrying out analysis and estimation of network conditions based on past packet measurements. Accordingly, they can make intelligent media adaptation decisions on behalf of the mobile devices. The cloud media proxy can fuse network measurement reports from many mobile users in the same coverage area, and derive robust statistical models for the wireless communication channel. In addition, as the cloud media proxy is typically situated half way along the path between sender and receiver, it can prompt the sender to take more agile adaptation actions in face of sudden changes, which is of particular importance for low-latency streaming applications. In this work, we showcase the potential benefits of cloud-assisted media streaming in the application scenario of mobile video confer- encing. For simplicity, we consider the option frame rate adaptation at the sender. The proposed cloud media proxy takes into consideration past observations such as measured round-trip-times and recent packet delivery delays, and dictates the mobile sender whether to encode or skip the next video content frame as captured by the camera. The intelligent frame skipping problem is formulated within the Markov decision process (MDP) framework. Our formulation captures the impact of many contributing factors to end-to-end system latency: time-varying nature of wireless communication channels, traffic shaping delay at the sender, and video content fluctuation. It is shown that the optimal frame skipping policy can be calculated using the stochastic dynamic programing (SDP) approach. Alternatively, the optimal solution can be closely approximated using a greedy heuristic scheme that only takes into account the most recent observed packet delivery delay. Simulation results show how system performance is influcence by various key factors, including application playout deadlines, network round-trip-times, and wireless link throughputs. The rest of the paper is organized as follows. The next section reviews related work. Section III provides an overview of the proposed cloud-assisted media streaming architecture. Section IV explains how we model the various components in end-to-end media streaming delay. Section V presents our MDP-based formulation of the intelligent frame skipping problem, together with the SDP-based optimal policy. In Section VI, we study fundamental performance tradeoffs of the system via simulation results under various network conditions. Section VII concludes the paper. II. RELATED WORK There exists a rich body of literature on low-latency video stream- ing in the conventional server-client architecture. For instance, the work in [3] has applied linear quadratic optimal control theory to the design of an optimal streaming rate controller, which achieves low startup delay, continuous playback, and efficient bandwidth utilization. In [4], the rate-distortion optimized packet scheduling problem is cast within the MDP framework, which accounts for random packet losses and delay over the best-effort network, as well as the interdependencies between media packets. The MDP approach is also used for video encoder rate control [5] and joint packet pruning and adaptive playout [6] over time-varying wireless channel. Our work follows a similar mathematical framework, but differs from the above by leveraging a cloud media proxy for calculating the media adaptation decisions on behalf of the mobile senders. Recent research has recognized the potential benefits of leveraging proxy servers at the cloud for augmenting the computational and power constraints of the mobile devices [7] [8]. It is shown in [9] that joint adaptation in rendering and encoded video qualities by the cloud gaming proxy can effectively improve the user experience. We, instead, consider generic low-latency media streaming applications in this work, and study their fundamental performance tradeoffs. International Conference on Computing, Networking and Communications Invited Position Paper Track 978-1-4673-0009-4/12/$26.00 ©2012 IEEE 949