Deep learning-based video analytics is computation-intensive. Manufacturers such as Nvidia have launched many embedded deep learning accelerators and are rapidly gaining market share. However, the computing resources of such accelerators are still limited and heterogeneous. Although existing systems aim at optimizing video query tasks from a variety of perspectives, they rarely consider the general cooperation between heterogeneous edge devices and the dynamic workload of video content. In this work, we present SplitStream, a distributed system for accelerating video query tasks across heterogeneous edge devices, which is able to fully utilize the resources on each device and adapt to the workload dynamics. The key to achieving this is the data parallelism brought by the multi-instance mechanism and the dynamic workload adaptability brought by the two-level workload balancing mechanism. Evaluation results show SplitStream reduces the result retrieval time by up to 19% and improves the resource utilization by up to 234%.

中文翻译：

---

SplitStream：边缘的分布式、工作负载自适应视频分析

基于深度学习的视频分析是计算密集型的。 Nvidia 等制造商推出了许多嵌入式深度学习加速器，并正在迅速获得市场份额。然而，此类加速器的计算资源仍然有限且异构。尽管现有系统旨在从多个角度优化视频查询任务，但很少考虑异构边缘设备之间的通用协作和视频内容的动态工作负载。在这项工作中，我们提出了 SplitStream，一种用于加速跨异构边缘设备的视频查询任务的分布式系统，它能够充分利用每个设备上的资源并适应工作负载动态。实现这一点的关键在于多实例机制带来的数据并行性和两级工作负载均衡机制带来的动态工作负载适应性。评估结果显示，SplitStream 可减少结果检索时间高达 19%，提高资源利用率高达 234%。