Optimization and Orchestration in the Cloud-to-Edge Computing Continuum

In this thesis, we improve existing orchestration techniques to address the new challenges the Cloud-to-Edge Computing Continuum raises.

Edge computing is a paradigm that moves computation and storage outside of the data centers to the edge of the network.
The Cloud-to-Edge Computing Continuum refers to the aggregation of computing resources from traditional data centers to the edge.
Edge Computing extends the capabilities of cloud computing.
Deploying servers close to end users reduces the delays and enables new use cases.
However, these geographically distributed machines create new challenges; they can be addressed by improving existing cloud computing techniques.
In this dissertation, we aim to simplify application deployment in the Cloud-to-Edge Computing Continuum.
We present three main contributions that help move toward that goal.

In the first part of this dissertation, we describe an experimental methodology to study orchestration in the Cloud-to-Edge Computing Continuum.
We evaluate the performance of a 5G core network deployed in the Computing Continuum to illustrate our methodology.

Then, we propose a new orchestration approach for reducing the costs of deploying applications in the Cloud-to-Edge Computing Continuum.
This orchestrator chooses an optimal location for deploying applications in terms of costs and quality of service.
It also offers a mechanism to update scheduling decisions when the environment changes.
We evaluate this new orchestration approach with a realistic 5G use case: Vehicular Cooperative Perception.

Finally, we study the performances of container CPU limitation mechanisms.
Setting limitations is essential to maximize the utilization of servers in the Computing Continuum, especially at the edge, where resources can be more limited.
However, the different CPU limitation mechanisms available offer different performances depending on the application type.
An inadequate setting could lead to negative impacts on the application's performance.
Therefore, we propose a methodology for automatically selecting the best CPU limitation mechanism.