Cloud Deployment Tradeoffs for the Analysis of Spatially Distributed Internet of Things Systems

Abstract

Internet-enabled devices operating in the physical world are increasingly integrated in modern distributed systems. We focus on systems where the dynamics of spatial distribution is crucial; in such cases, devices may need to carry out complex computations (e.g., analyses) to check satisfaction of spatial requirements. The requirements are partly global—as the overall system should achieve certain goals—and partly individual, as each entity may have different goals. Assurance may be achieved by keeping a model of the system at runtime, monitoring events that lead to changes in the spatial environment, and performing requirements analysis. However, computationally intensive runtime spatial analysis cannot be supported by resource-constrained devices and may be offloaded to the cloud. In such a scenario, multiple challenges arise regarding resource allocation, cost, performance, among other dimensions. In particular, when the workload is unknown at the system’s design time, it may be difficult to guarantee application-service-level agreements, e.g., on response times. To address and reason on these challenges, we first instantiate complex computations as microservices and integrate them to an IoT-cloud architecture. Then, we propose alternative cloud deployments for such an architecture—based on virtual machines, containers, and the recent Functions-as-a-Service paradigm. Finally, we assess the feasibility and tradeoffs of the different deployments in terms of scalability, performance, cost, resource utilization, and more. We adopt a workload scenario from a known dataset of taxis roaming in Beijing, and we derive other workloads to represent unexpected request peaks and troughs. The approach may be replicated in the design process of similar classes of spatially distributed IoT systems.