Node Manager

Edge OS

The Edge OS is a Linux distribution customized with Yocto to include only the software components required on FPGA-based edge nodes, for orchestration and for the management of FPGA accelerators. It is part of a unified design flow for FPGA-accelerated workloads on edge MPSoCs, which covers the development of the FPGA accelerators, the containerization of the workloads, and their deployment through lightweight Kubernetes (K3s). It supports both Vitis AI-based DPU accelerators and custom FPGA IPs, providing the software, drivers and run-time components needed to use them from containerized workloads.

Node Manager

The Node Manager provides local run-time management of heterogeneous edge nodes, such as the HMPSoC accelerators and the RISC-V-based processor with CGRA extensions. The CPU hosts the operating system and the Kubernetes environment that integrate the device into the computing continuum, while the FPGA accelerates the computationally intensive parts of the workload. Two complementary mechanisms keep the node in an efficient operating state.

Energy management. The OS-based manager continuously monitors hardware resources and workload execution, analyzes system conditions, selects the most suitable operating configuration, and dynamically applies hardware-level reconfigurations: CPU frequency scaling, CPU core activation, and workload execution on CPU or FPGA. It follows a MAPE-K autonomic control loop (Monitor, Analyze, Plan, Execute, with shared Knowledge), distinguishes active workloads from idle conditions, and supports containerized applications without requiring application-specific modifications.

Accelerator scheduling. Power and performance traces collected while the workload runs train machine-learning models at run time (run-time workload characterization). A scheduler uses their predictions whenever the workload changes to decide which pending acceleration tasks to deploy and how many accelerator replicas to assign to each, favouring configurations with less interference between accelerators.

Edge OS

The design flow and the Edge OS have been validated on AMD Kria KV260 and Digilent ZedBoard platforms, with both AI-based and custom hardware workloads. Compared with vendor-provided Linux distributions, the environment strongly reduces memory usage, flashing time, idle CPU usage and container sizes, while workload execution time and power and energy consumption remain broadly comparable. The Edge OS is also used as the infrastructure support for the evaluation of the Node Manager energy management.

Node Manager

Within MYRTUS, the OS-based manager has been developed with native support for heterogeneous CPU–FPGA platforms; a monitoring subsystem collecting hardware- and workload-level metrics (CPU utilization, memory, power, operating frequencies, I/O, network traffic and container resource utilization); a closed-loop architecture that selects the most appropriate operating configuration according to energy and performance objectives; dynamic hardware reconfiguration; and predictive models that can be validated and updated at run time. On an AMD Kria KV260 platform, the energy savings grow as the adaptation includes frequency scaling, then core activation, then FPGA offloading; the run-time detection of idle conditions is being evaluated within the complete control loop.

The scheduler explores a limited set of candidate configurations with a lightweight metaheuristic (the Crow Search Algorithm), and a configurable weighting factor balances performance against a fair sharing of the FPGA resources. On AMD Zynq UltraScale+ platforms with dynamic workloads, it reduced the workload makespan and the energy consumption of CPU and reconfigurable fabric compared with first-come, first-served scheduling, and the same configuration worked on platforms with very different resources without retuning.