A variety of use cases of edge computing, including automation, augmented reality, autonomous driving require predictable communication and compute systems as well as application responsiveness due to the tight integration of the applications with the physical world.
Challenges and goals
The requirements of predictable communication and compute systems must be met despite the openness of edge computing systems, the usage of off-the-shelf components as well as the shared nature of the edge infrastructure.
The predictability project targets fundamental contributions that enable predictable edge computing infrastructures and services. It investigates how predictability is best combined at runtime with relevant applications and what the consequences are for the management of large edge systems.
Tasks and Methodologies
The project will focus on three tightly coupled objectives.
Data-driven Performance Modeling and Optimization
The first objective of this project is to is to develop tractable, data-driven models of end-to-end system performance. We are interested in fundamental questions, such as how different wireless transmission technologies translate into different latency and reliability profiles, how corresponding computing technologies and abstractions (operating systems, virtualization, containers) can be made predictable, and how the concatenation of communication and compute functions translates into stochastic latency and reliability profiles.
These profiles serve as input to optimization where compute resources can be leveraged to compensate for communication deficiencies, and vice versa. We will approach these steps through performance modeling and optimization theory on the one hand, and we envision to approach the parameterization of fundamental models through machine learning. We will investigate how theoretical results can translate into practically relevant results, and how this depends on the considered systems and environments.
The second objective is to develop models and tools for providing predictability for edge computing systems in a larger infrastructure context. A key aspect of system wide predictability is the interaction between the network core and the edge systems, for instance with respect to orchestration and network slicing.
Further challenges arise from automating resource reservation and task allocation in large edge infrastructures, for instance with respect to load balancing, life-cycle management, detection and mitigation of latency/reliability anomalies while preserving the predictability properties for the running applications. Systems theory, optimization theory, as well as game theory will be the main means to address these issues.
Adaptive System Design
The third objective is to ensure predictability at the application layer. The the first fundamental question to address is what a certain latency/reliability profile implies for a running application, and how application requirements can be met and optimized with the infrastructure profiles. We will address stateful runtime management of edge systems and applications, e.g., for application migration due to mobility or for load balancing, as well as actively steering the latency/reliability requirements of applications such as collaborative CPS.
A second fundamental question to address, in particular for collaborative CPS, is the ability to ensure redundancy and the potential of anytime algorithms to deal with rare events in the latency/reliability profiles. To address these we will resort to tools from control theory, hybrid systems theory, as well as scheduling theory.
Edge computing, Resource management
Mechatronics, Supervisory ctrl, Optimization, Model-checking
Machine Intelligence, Telecom and distributed systems, R&D, Leadership and management
Wireless, Predictability, Edge computing