The goals of CloudSkin are the following:

• Smart management for the Cloud-edge continuum: The overall objective is to leverage the generated knowledge from state-of-art AI methods to transparently orchestrate Cloud-edge resources. The key goal is to build a “Learning Plane” that, in cooperation with the application execution framework and continuum infrastructure, can enhance the overall orchestration of Cloud-edge resources. Such plane is the materialization of the cognitive cloud, where decisions on the cloud and the edge are driven by the continuously obtained knowledge and awareness of the computing environment through AI, and particularly, neural networks and statistical learning, taking the challenge of enabling these methods into low-power edge devices.

• Virtual execution for the Cloud-edge continuum: This goal focuses on a new universal and flexible execution abstraction, we called it “Cloud-edge cells”, that will enable the execution of legacy and highly granular applications in the cloud continuum. The new container-like execution abstraction will be based on the WebAssembly technology. It will enable the execution of the same computation on a wide range of cloud and embedded devices and make task execution migratable across different servers and devices in the continuum infrastructure. We will integrate our WebAssembly executor with Kubernetes. More specifically, we will contribute new features to Kubernetes that will support the efficient migration of WebAssembly containers between different levels of the continuum, exploiting WebAssembly’s capability for state serialization.

• Infrastructure support for the Cloud-edge continuum: This objective is to prepare the infrastructure to turn it into a virtual resource continuum, where the large set of Cloud-edge cells composing applications can be allocated flexible resources, according to their dynamically changing needs. One of the major challenges here is to design an infrastructure to support extremely short-lived Cloud-edge cells and tasks (of 1 to 10ms, or less) and extremely intense bursts with fast data access requirements. This requires delivering bare metal resource performance to storage, despite virtualization and dynamic reallocation, which today is not possible in the cloud continuum. CLOUDSKIN will achieve this by leveraging high-performance I/O (RDMA networking) and near-storage CPU compute capacity (GPUs, FPGAs) to the fine-grained application tasks.