QoS Aware Video Analysis Over Low-Cost Edge-Cluster: A Utility Minimization Approach

The constrained availability of resources on an edge analytics platform prompted the need for a trade-off between accuracy and latency by selecting suitable deep neural network (DNN) models on-the-fly. Earlier efforts either used a single powerful multi-core edge computing device or a distributed cluster of edge nodes. While the former has a high cost and power consumption, the latter incur a high communication overhead. In this paper, we propose a quality-of-service (QoS) aware video analytics platform using an edge-cluster made of low-cost devices. The edge nodes, that constitute the cluster, host heterogeneous DNN models having different configurations and number of layers. The nodes cooperate among themselves to jointly process a streaming video to achieve an optimal QoS. We formulate an optimization problem using penalty as the utility function to minimize the long-term average penalty (LTAP). We first design a DNN model recommender algorithm to minimize the LTAP and then compare it with an Oracle to show that it can achieve an LTAP with an error of 1.6 % and 9.88 % for video resolutions of 720p and 2160p respectively. We also show that the bounds on LTAP are lower and tighter for lower resolution videos compared to the higher resolution videos.