Hybrid Scheduling for Deep Learning Accelerators

The patent combines multiple scheduling methods (reinforcement learning-based scheduling and heuristic-based scheduling) into a hybrid scheduling system. The reinforcement learning component handles complex optimization for energy consumption and resource efficiency, while the heuristic component provides fast, rule-based scheduling for immediate task allocation. This merging allows the system to achieve high resource efficiency without being overwhelmed by a single complex method.

The scheduling system is segmented into multiple independent scheduling modules, each handling specific aspects of resource allocation. The reinforcement learning module focuses on long-term optimization and energy efficiency, while heuristic modules handle specific scheduling rules and constraints. This segmentation allows the system to manage complexity by dividing the scheduling problem into smaller, manageable parts that can be optimized independently.

The system applies partial optimization by using reinforcement learning for high-level resource allocation decisions that have significant energy impact, while using faster heuristic methods for detailed scheduling adjustments. This partial application of complex optimization only where most needed allows the system to reduce overall energy consumption without incurring the full computational overhead of optimizing every scheduling detail through reinforcement learning.

The scheduling system dynamically switches between different scheduling strategies based on current system conditions. When energy optimization is critical and system load is manageable, the system uses reinforcement learning-based scheduling. When speed is critical or system state changes rapidly, the system switches to faster heuristic-based scheduling. This dynamic adaptation allows the system to optimize energy consumption without being permanently constrained by slow scheduling times.