Markov Blanket Composition of SLOs

Abstract

Smart environments use composable microservices pipelines to process Internet of Things (IoT) data, where each service is dependent on the outcome of its predecessor. To ensure Quality of Service (QoS), individual services must fulfill Service Level Objectives (SLOs); however, SLO fulfillment is dependent on resources (e.g., processing or storage), which are scarcely available within the Edge. Hence, when distributing services over heterogeneous devices, this raises the question of where to deploy each service to best fulfill both its own SLOs as well as those imposed by dependent services. In this paper, we maximize SLO fulfillment of a pipeline-based application by analyzing these dependencies. To achieve this, services and hosting devices alike are extended with a Markov blanket (MB) - a probabilistic view into their internal processes - which are composed into one overarching model. Given a mutable set of services, hosts, and SLOs, the composed MB allows inferring the optimal assignment between services and edge devices. We evaluated our method for a smart city scenario, which assigned pipelined services (e.g., video processing) under constraints from subsequent services (e.g., consumer latency). The results showed how our method can support infrastructure providers by optimizing SLO fulfillment for arbitrary devices currently available.