SoCRATES: System-on-Chip Resource Adaptive Scheduling using Deep Reinforcement Learning

Deep Reinforcement Learning (DRL) is being increasingly applied to the problem of resource allocation for emerging System-on-Chip (SoC) applications, and has shown remarkable promises. In this paper, we introduce SoCRATES (SoC Resource AdapTivE Scheduler), an extremely efficient DRL-based SoC scheduler which maps a wide range of hierarchical jobs to heterogeneous resources within SoC using the Eclectic Interaction Matching (EIM) technique. It is noted that the majority of SoC resource management approaches have been targeting makespan minimization with fixed number of jobs in the system. In contrast, SoCRATES aims at minimizing average latency in a steady-state condition while assigning tasks in the ready queue to heterogeneous resources (processing elements). We first show that existing DRL-based schedulers developed with the makespan minimization objective are ineffective for the latency-minimization-driven SoC applications due to their characteristics such as high-frequency job workload and distributed/parallel job execution. We then demonstrate that through its EIM technique, SoCRATES successfully addresses the challenge of concurrent observations caused by the task dependency inherent in the latency minimization objective. Extensive tests show that SoCRATES outperforms other existing neural and non-neural schedulers with as high as 38% gain in latency reduction under a variety of job types, queue length, and incoming rates. The resulting model is also compact in size and has very favorable energy consumption behaviors, making it highly practical for deployment in future SoC systems with built-in neural accelerator.