Pervasive applications over large-scale, distributed embedded devices and the Internet of Things (IoT) demand precise coordination with the network; for example, several such applications, like collaborative video streaming and live analysis, augmented reality, etc., need continuous monitoring of network throughput and adapt the application behavior accordingly. Although the idea of network throughput prediction is not new and quite dated, in this paper, we show that the existing approaches fail to correctly infer the throughput when the network operator or the device change, and thus, not generic enough for Internet-scale applications. We propose \ourmethod, a novel approach that allows collaborative training across different client hardware by capturing throughput variations based on devices' sensitivity towards the corresponding network configurations. Rigorous evaluations show that \ourmethod{} outperforms various standard baseline algorithms with more than $80\%$ R2-score over different datasets. We also analyze the performance of \ourmethod{} over a network-aware streaming media application and demonstrate its efficacy for various application scenarios.

翻译：大规模分布式嵌入式设备与物联网中的泛在应用，需要与网络进行精确协调；例如协同视频流与实时分析、增强现实等多种应用，需要持续监控网络吞吐量并相应调整应用行为。尽管网络吞吐量预测的概念由来已久，但本文表明现有方法在网络运营商或设备变更时无法正确推断吞吐量，因此难以满足互联网级应用的通用性需求。我们提出\ourmethod这一新型方法，通过捕捉设备对相应网络配置的敏感性差异，实现跨不同客户端硬件的协作训练。严格评估表明，\ourmethod在不同数据集上的R2分数均超过80%，显著优于多种标准基线算法。我们还分析了该方法在网络感知流媒体应用中的性能，并论证了其在多种应用场景下的有效性。