A Multi-agent Approach to Self-Distributing Systems

Abstract

As the computing continuum increasingly becomes the default deploy- ment infrastructure for modern systems, the demand for a programming model capa- ble of meeting the requirements for developing software systems on such infrastruc- tures also grows. This model must provide abstractions to manage the high level of dynamism inherent in these environments, particularly addressing the autonomous management of stateful service placement across large-scale edge-cloud continuum infrastructures. To address this issue, we explore the concept of self-distributing systems – a machine-centric approach to deal with the complexity of designing dis- tributed systems. We take a step further and have the system decide on distributed design decisions at runtime as unexpected changes and events occur, leaving the sys- tem responsible for reacting quickly and accurately as a response to such changes. This paper presents the application of a multi-agent learning approach to learn how to distribute stateful services across the continuum. We demonstrate the efficiency of such a method in a local testbed. We compare our results against a multi-armed bandits approach, pinpointing the strengths and weaknesses of the two approaches.