Emerging applications of edge to cloud compute continuum with faster connectivity include collaborative intelligence, automated driving, and augmented/virtual reality. Here, the computationally intensive tasks are to be offloaded to the edge compute servers installed at the multi-vendor telco edge. This allows devices to remain light weight and conserve battery.
In the network edge, computing and storage resources are typically distributed across communication service provider (CSP) premises, between national, regional and local access sites (https://www.ericsson.com/en/edge-computing). In future, the radio base stations could also be provisioned with additional compute capacity to support ultra-low latency applications. Devices will request different computation tasks requiring some specified number of resources to be executed on any suitable edge server. Meanwhile, each device will be within range of multiple edge servers. The challenge is determining which device connects to which edge server to maximise the number of UEs served, while ensuring the resource demands of allocated UEs do not exceed the capacity of the serving edge server. This problem is called the Edge User Allocation (EUA) problem. The assignment can be dynamically adapted based on UEs location, and resource availability. Existing techniques for solving the EUA problem are expensive in terms of compute resource efficiency, struggle to scale with the problem size and trade off time to solution with its quality.
To circumvent these limitations, the EXTRA-BRAIN use case lead by Ericsson adopts the hardware/software co-design approach where brain inspired neural networks (SNN, BCPNN, etc.) are tailored made and accelerated on the Field-Programmable Gate Array (FPGA). For example, the neurons in the BLNN represents the possible UE-to-server assignments, and the connections among the neurons encode the constraints. The neurons are excited or inhibited at runtime by monitoring the quality of assignments and adjusting the excitation/inhibition to yield acceptable quality solution. By accelerating the BLNN representing the largest problem instance on FPGA and configuring the active processing elements at runtime can essentially lead to the neuromorphic solver capable of solving EUA type problems also in the dynamic setting.
Below the plausible schematic of FPGA accelerated neuromorphic solver accessible via Kubernetes engine is shown. Depending on the current snapshot of the network, expressed as specifications of EUA problem instance, the device plugin could reconfigure the BLNN accelerator and generate the near-optimal UE-edge server assignments. Kubernetes could then orchestrate the UE specific WebAssemby workloads on the assigned edge server nodes.
