Speaker
Description
We present the initial design and proposed implementation for a series of long-baseline, distributed inference experiments leveraging ARA --- a platform for advanced wireless research that spans approximately 500 square kilometers near Iowa State University, including campus, the City of Ames, local research and producer farms, and neighboring rural communities in central Iowa. These experiments aim to demonstrate, characterize, and evaluate the use of distributed inference for computer vision tasks in rural and remote regions where high-capacity, low-latency wireless broadband access and backhaul networks enable edge computing devices and sensors in the field to offload compute-intensive workloads to cloud and high-performance computing systems embedded throughout the edge-to-cloud continuum. In each experiment, a distributed implementation of the MLPerf Inference benchmarks for image classification and object detection will measure standard inference performance metrics for an ARA subsystem configuration under different workload scenarios. Real-time network and weather conditions will also be monitored throughout each experiment to evaluate their impact on inference performance. Here, we highlight the role of HTCondor as the common scheduler and workload manager used to distribute the inference workload across ARA and beyond. We also discuss some of the unique challenges in deploying HTCondor on ARA and provide an update on the current status of the project.
