A Hierarchical Neural Task Scheduling Algorithm in the Operating System of Neuromorphic Computers

Bionic computing, increasingly favored for its sophisticated approach to knowledge applications, is experiencing a revolution bolstered by neuromorphic hardware, which delivers versatile solutions in a multitude of scenarios. In this context, we introduce DarwinOS Scheduler-a hierarchical distributed operating system framework optimized for configurable multi-core neuromorphic chips with synchronous communication. This scheduler adeptly manages neuronal computation tasks influenced by data streams, promoting dynamic spiking neural network (SNN) operations where process states switch responsively to activity events. To bolster efficiency, DarwinOS incorporates general-purpose executors for task-related data handling, both prior to and after core processing. Alongside, we propose a neuromorphic task scheduling method, Hierarchical Distributed Scheduling for Neuromorphic tasks (HDSN), that dynamically identifies and leverages common computational patterns among various tasks. This maximizes data processing while adhering to hardware constraints. Simulations affirm that DarwinOS scheduler with HDSN outperforms traditional methods, boosting system throughput and resource utilization by 12% and 4%, respectively, thus enhancing the performance and efficiency of large-scale neuromorphic systems.