Heterogeneous Embedded Resource Management under Space-based Edge Computing Environment with Kubernetes Device Plugin

Abstract

Edge devices are widely applied in space scenarios for their compact size and diminished power consumption. To avoid the collision of different applications, containerizing applications becomes a more welcomed option compared to virtual machines for its light-weightiness and short loading latency. Kubernetes, considered a de facto framework for orchestrating containers, allows automatic allocations of resources connected to nodes, but it only supports management on CPU, I/O, network, and storage by default. Supported devices can be extended with its device plugin mechanism however, official device plugins might not work on edge devices, as their functionalities are constrained due to its compactness, such as NVidia TX2 that is not supported by the official Nvidia GPU device plugin. Moreover, devices vary in type, and architecture and even some do not support operating systems, making it unintelligent to develop device plugins separately. Therefore, managing device plugins becomes an urgent problem for edge cloud providers. In this paper, we proposed and implemented a paradigm of orchestrating devices that allows Kubernetes manage edge devices, even if they do not run operating systems. As for the obstacles of being unable to report the device topology of a single node to the scheduler, we proposed an approach to hijack the scheduling process without establishing a new scheduler. We conducted experiments and the results showed that the platform we proposed can enable Ku-bernetes to manage various devices without device plugins, both on on-board embedded devices and non-OS-supported devices. We tested its performance and capacity with a simulated device, and the result shows that it only causes 6.34 % overhead, and its embedded scheduling method has an adequate performance compared to optimum method, which is only 7.7 % slower.