Causality-Based Service Adaptations for Elastic Computing Continuum Systems

Abstract

Computing Continuum (CC) systems are challenged to ensure the intricate requirements of multiple computational tiers. While some tiers, like the Cloud, can easily scale their processing resources to meet specific Service Level Objectives (SLOs), other tiers, like Edge computing, lack this abundance of resources and must adhere to alternative ways for elasticity. To that extent, creating a causal view into processing services allows to accurately infer the best possible way to adjust a service, and in further consequence, scale composite services distributed throughout the CC. We identify three challenges that must be tackled to achieve this: (1) to identify key factors that drive SLO fulfillment, it needs a transparent view into individual services and their precise implications to the processing hardware; given that, the next step is to (2) estimate the impact that processing services have on each other’s SLOs, which is defined by the interactions and dependencies of services; lastly, (3) to ensure that services choose the best adaptations regardless of external perturbations, any causal model must be continuously adjusted according to new observations. This thesis proposes novel methods to achieve overarching SLO fulfillment in CC systems, which use Active Inference (AIF) to create a causal view of individual services as well as their interactions. Thus, the CC can be spanned with SLOs of different granularity, so that individual services, or compositions of services, can autonomously achieve an equilibrium.