Action-Based Scheduling: Leveraging App Interactivity for Scheduler Efficiency
Action-Based Scheduling: Leveraging App Interactivity for Scheduler Efficiency
复制标题
DOI：
10.1109/tnet.2018.2882557
复制
发表时间：
2019-02
期刊：
影响因子：
--
通讯作者：
John Tadrous;A. Eryilmaz;A. Sabharwal
中科院分区：
文献类型：
--
作者：
John Tadrous;A. Eryilmaz;A. Sabharwal
智能手机流量的主要部分来自涉及人类互动的应用程序。特别是，当人类用户从服务器接收信息时，他们在采取行动之前花费几秒钟的信息处理时间。用户处理时间在应用程序会话期间创建空闲通信周期。此外，未来流量的生成取决于当前查询-响应对的服务。在本文中，我们的目标是利用这种交互的特性来获得体验质量收益。现有的调度器在实践和理论上都不是考虑到上述业务特性而设计的。理论工作主要集中在独立生成或直接控制的流量调度上，但不受人类交互引起的特定动态的支配。另一方面，在实践中，调度器使用轮询和处理器共享方法来服务多个正在进行的会话。我们表明，这两种方法对于提供涉及人工交互的应用程序都不是有效的。相反，我们展示了交互流量的最优调度在分组上是非随机化的，我们称之为基于动作，因为它避免中断正在进行的动作服务，以便使人类响应时间与其他动作的服务保持一致。由于基于行动的最优策略的设计在计算上是不可行的，我们开发了低复杂度的次优基于行动的策略，该策略对于正在进行的两个会话是最优的。基于真实数据跟踪的数值研究表明，与基于分组的处理器平等共享相比，我们提出的基于动作的策略可以将总时延降低22%。
The dominant portion of smartphone traffic is generated by apps that involve human interactivity. Particularly, when human users receive information from a server, they spend a few seconds of information processing before taking an action. The user processing time creates an idle communication period during the app session. Moreover, the generation of the future traffic depends on the service of the current query-response pair. In this paper, we aim at leveraging the properties of such interactions to reap quality-of-experience gains. Existing schedulers, both in practice and theory, are not designed in view of the aforementioned traffic characteristics. Theoretical works predominantly focus on scheduling of traffic that is either generated independently or directly controlled, but not governed by the specific dynamics caused by human interactions. Schedulers in practice, on the other hand, employ round-robin and processor-sharing methods to serve multiple ongoing sessions. We show that neither of these approaches is effective for serving apps that involve human interactivity. Instead, we show that optimal scheduling for interactive traffic is non-randomized over packets, which we call action-based, as it avoids breaking ongoing service of actions in order to align human response times with the service of other actions. Since the design of optimal action-based policy is computationally prohibitive, we develop low-complexity suboptimal action-based policies that are optimal for two ongoing sessions. Our numerical studies based on a real-data trace reveal that our proposed action-based policies can reduce total delay by 22% with respect to packet-based equal processor sharing.