Linearize, predict and place

Abstract

Many IoT applications found in cyber-physical systems, such as smart grids, must take control actions in response to critical events, such as supply-demand mismatch, which requires low-latency processing of streaming data for rapid event detection and anomaly remediation. These streaming applications generally take the form of directed acyclic graphs (DAGs), where vertices represent operators and edges represent the flow of data between these operators. Edge computing has recently attracted significant attention as a means to readily meet the requirements of latency-critical IoT applications due to its ability to provide low-latency processing near the source of data. To accrue the benefits of edge computing, the constituent operators of these applications must be placed in a manner that intelligently trades-off inter-operator communication costs with the cost of interference incurred due to co-location of operators on the same resource-constrained edge devices. To address these challenges and to substantially simplify the placement problem for DAGs of arbitrary sizes and topologies, we present an algorithm that first transforms any arbitrary stream processing DAG into an approximate set of linear chains. Subsequently, a data-driven latency prediction model for co-located linear chains is used to inform the placement of operators such that the makespan, defined as the maximum latency of all paths in the DAG, is minimized. We empirically evaluate our algorithm using a variety of DAG placement scenarios on a Beagle Bone cluster, which is representative of an edge computing environment.