Control plane for situation-awareness applications on geo-distributed resources

Abstract

Situation-awareness applications generate actionable knowledge from sensor and user data. Two trends are unlocking new situation-awareness applications: geo-distributed resources and pervasive sensors. Geo-distributed computing infrastructure is now available worldwide, ranging from multi-region cloud deployments to newer 5G deployments. On the other hand, pervasive sensors have seen a boom, with examples like geo-distributed camera deployments, smart cities, and users’ gadgets (smartphones). Having computational resources closer to the data source improves the response time and the efficient use of the available resources. Geo-distributed resources reduce the physical distance to the data source, cutting the time it takes to transmit, filter, and process information, reducing unnecessary data transmission. However, efficient management of resources is challenging for densely geo-distributed resources while also providing spatio-temporal context and latency quality-of-service objectives.

This dissertation proposes a control plane that makes three contributions for efficiently managing situation awareness applications running on geo-distributed computational resources: 1. It defines a programming model capable of expressing situation-awareness applications and their requirements. Additionally, it defines a new taxonomy for situation-awareness applications. 2. It describes the requirements of the control plane and the components needed to support geo-distributed resources and situation-awareness applications. 3. It proposes an efficient control plane architecture and mechanisms to support the above requirements.

First, this work extends the data-flow graph programming model to better suit the geo-distributed context of both applications and computational resources. Then, we analyze and classify distinct types of situation-awareness applications that would benefit from geo-distributed resources. Finally, it presents ways to represent the requirements of such applications as part of an extended data-flow graph.

Second, this thesis defines the different building blocks required to manage geo-distributed resources efficiently. Then, starting from the programming model, we show how the different components interact with the geo-distributed physical computational resources. Finally, we analyze why state-of-the-art centralized control planes cannot fulfill situation-awareness application’s requirements.

Third, this dissertation contributes Foglets, a fully decentralized control plane architecture. A fully decentralized design reduces the overhead of a centralized design while still providing low-latency access to computational resources. In addition, we present mechanisms to propagate information to allow entirely local decisions and support application migrations between different computational resources’ localities.

Finally, this thesis proposes OneEdge, a hybrid control plane architecture, extending the decentralized and centralized designs from the previous section. The control plane for geo-distributed resources has an inherent trade-off between response time and the quality of decisions. The main factor that defines the trade-off is the transmission latency incurred between the resources being managed and the control plane component making those decisions for the resources. For example, if the control plane decisions are made in a centralized fashion, we can calculate optimal decisions, but there is a high likelihood that all these decisions would incur at least one wide-area network round-trip, given that the resources are geo-distributed. On the other hand, the latency can be minimized if the control plane components are close to the resources, but then the complexity and cost of maintaining an up-to-date view of other resources and making optimal decisions are increased. We present an architecture that allows us to configure the control plane for different trade-off points and provide all the required service-level objectives presented in the programming model chapter.