Node Configuration and Computational Capabilities

Our fixed and portable site equipment will include modular and configurable nodes that can be customized per use case and experiment objectives. A unique feature is local processing power at using low-power Graphics and Tensor Processing Units (GPU/TPU) to enable a range of application needs. The TPUs will be supported by the Tensorflow Lite Library and the PyCoral (Python), Libcoral C++), and Libedgetpu (C++) API’s, enabling rapid development and deployment of CNNs for image/video processing and general classification applications. Some experimenters will prefer to use GPUs; we will make available plug-in NVIDIA Jetson 4GB Nano processors to accelerate learning algorithms, e.g., for video-based applications. Exemplary applications that utilize GPU/TPU powered nodes to enable distributed computing, federated learning, on-the-fly compression for UAV networks, as well as applications to fire growth modeling. This architecture will also provide experimenters with the flexibility to evaluate distributed learning algorithms and data sharing policies; this includes applications employing ground and aerial robotic nodes with sensing and actuation capabilities. As used for the successful IoT-Lab, the DISCOVER nodes consist of two major components: a programmable experimental node that is exposed to experimenters, and a supervisor node that facilitates remote programming and monitoring of the experimental node. The experimental node will be based on the Raspberry Pi 4 Computer Module and I/O Board with a LoRa shield and optional shields for FPGA-based computing, as well as optional 4 GB Jetson Nano or google Coral TPU accelerators. The supervisor node will be based on a Raspberry Pi 4 single-board computer. This plan is tentative; we will re-evaluate in the context of rapidly-changing possibilities to maximize performance and flexibility if the project receives an award. The experimental node board plugs into the supervisor node, enabling both comprehensive control and monitoring of experiments and the flexibility of adding the CISE researchers’ custom-designed boards using standard interfaces. The supervisor node receives program images for the experimental node from DISCOVER servers via WiFi 802.15.4, or 4G/5G and communicates with the experimental node’s bootloader via USB for downloading and verifying program images. We will also provide to experimenters monitoring and control functions of the experimental node as well as configurability of the supervisor node via a set of parameters and an open API. For security, experimenters will not be able to program the supervisor node; if necessary for a particular experiment, the DISCOVER team will generate and deploy custom supervisor node software. Use of a Linux-based SBC for the supervisor node will enable us to leverage the vast open-source space for both experimenters and the project team.