Characterizing the Performance of Distributed Edge Processing Resource Allocation in Dynamic Networked Environments

Distributed edge processing in resource-constrained networks relies on accurate understanding of available resources to enable effective deployment of complex workflows to execute analytics. This paper presents an initial analysis characterizing the impact of measurement error on application placement and performance of distributed analytics. We conduct experiments on a distributed edge processing cycle including resource monitoring, resource allocation, and analytics execution phases. To facilitate experimentation of situations with a broad range of monitoring error, we leveraged a machine learning-based model trained on empirically obtained performance metrics to predict analytics performance. We also introduce error into the resource monitoring steps, which result in diminished performance and different resource placement dynamics. Our initial results suggest that more accurate/frequent CPU metrics reporting should be prioritized in this system, rather than bandwidth metrics.