Holistic Allocation of Edge and Wireless Resources for Time-Sensitive IoT Applications

Resource provisioning and scheduling for edge and cloud architectures have gained a lot of attention in the recent past both in academia and industry. This interest is primarily driven by the enormous success of two technologies in the field of computing: 1) Artificial Intelligence (AI), and 2) Internet of Things (IoT). Cloud architectures play a central enabling role for AI due to the significant demand that this technology imposes on computational resources. Orthogonally, IoT has enabled an unprecedented decentralization and deployment of applications across domains, including several in which AI (at the edge) is gaining prominence. As a result, the edge and cloud hardware architecture is poised to dominate the field of computing across application domains in the future. The general objective of this project is to minimize the power consumption of end devices by offloading their computation intensive workload on edge and cloud servers. Formulated as such, this problem has already received a lot of attention in the literature on mobile and edge computing. However, the originality of our work lies in the consideration of timing constraints associated with the workload: in building cooling systems for example, some of the software services such as real-time control are indeed time-sensitive which means they must be performed within predefined deadlines. This aspect has also been studied in a few research works, but most of them consider simplistic hypothesis when it comes to modelling timing interference among workload. The problem we propose to address is thus a resource allocation problem with the following objectives: (i) To minimize power consumption of end devices; (ii) To minimize the average response times of non time-sensitive workload; and (iii) To enforce end-to-end deadlines of time-sensitive workload with guarantees. As opposed to existing works, which study a subset of the infrastructure, we aim at a holistic approach considering (i) the wireless communication resources, (ii) the computation resources of edge and cloud servers and end devices, as well as (iii) the communication resources of the backhaul network. This problem is ambitious and will require finding a trade-off between optimality and reactivity: on the one hand, resource allocation problems are usually difficult to solve optimally because of their combinatorial complexity. On the other hand, service workload characteristics vary over time (notably because of mobility of devices), which means the resource allocation framework must run online together with the system in order to adapt resource allocations to these variations. In this project, we aim to address this complex resource allocation problem using a decentralized divide-to-conquer strategy. The infrastructure will be divided into segments depending on the architecture of the backhaul network, with each segment comprising a set of potentially different resources. Resources will be independently provisioned (mapping of workload to computation and communication resources) and scheduled (allocation of a resource to workload over time) within each segment. We plan to develop consensus based decentralized provisioning algorithms for this purpose, and aim to achieve convergence as well as bounded approximation factors using submodular objective functions. For the scheduling of computation resources and the backhaul network with deadline guarantees, we plan to develop delay composition techniques along with priority assignment strategies based on extensions to existing approaches from the distributed real-time embedded systems area. On the other hand, for the scheduling of wireless network resources, we plan to focus on a specific protocol, LoRa, and develop spectrum and duty cycle allocation algorithms that can provide probabilistic guarantees on the communication delay. An iterative strategy will also be designed to decompose the end-to-end deadlines of time-sensitive workload into segment level deadlines, such that response times of non time-sensitive workload can be minimized while ensuring the deadline feasibility of time-sensitive workload. Finally, to demonstrate the developed capabilities, we also plan to implement these resource allocation algorithms in a urban cooling system testbed in NTU.