Perception Workload Characterization and Prediction on the Edges with Memory Contention for Connected Autonomous Vehicles

Vehicular Edge computing requires computational power from connected Edge devices in the network to process incoming vehicle work requests. This connection and offloading allows for faster and more efficient data processing and thus improves the safety, performance, and reliability of the connected vehicles. Existing works focus on the processor and its characterization, but they forgo the connecting components. Memory resource and storage resource is limited on Edge devices, and the two combined incur a heavy impact on deep learning. This is prominent as perception-based workloads have yet to be studied deeply. In our characterization, we have found that memory contention can be split into 3 behaviors. Each of these behaviors interacts with the other resources differently. Then, in our deep neural network (DNN) layer analysis, we find several layers that see computation time increases of over 2849% for convolutional layers and 1173.34% for activation layers. Through the characterization, we can model the workload behavior for the Edge based on the device configuration and the workload requirements. Through this, the impacts of memory contention and its impacts are quantified. To the best of our knowledge, this is the first such work that characterizes the memory impacts towards vehicular edge computational workloads with a deep focus on memory and DNN layers.