A Comparison of Static and Dynamic Micro-Service Placement Strategies for Edge Computing

The advent of edge computing leverages large, distributed infrastructures of edge nodes for executing complex applications. These applications are typically based on the composition of multiple micro-services. The problem of optimal allocation of micro-services to the infrastructure is a typical problem that aims to reduce power consumption and response time. However, as the workload changes, a trade-off arises between the need to define new deployments to support the changed system conditions and avoiding complex re-configurations of the infrastructure that may negatively impact the system availability (due to the need to switch on and off nodes and due to the number micro-services migrations). We propose an optimization model based on dynamic programming that is explicitly designed to reduce the disruption on the edge node infrastructure by limiting the number of reconfiguration operations while still guaranteeing the respect of Service Level Agreements in terms of average response time for the various applications. Our model considers the co-existence of micro-services belonging to different applications on the same nodes as well as the impact of network delays. A comparison with a naive alternative that defines a new deployment every time that the workload changes demonstrates that the proposed dynamic programming approach can reduce the number of infrastructure re-configurations by at least 75% and the number of nodes switched on/off by 72%, while still guaranteeing that the power consumption is close to the minimum and response time can still satisfy the SLA requirements.