CuSH: Cognitive Scheduler for Heterogeneous High Performance Computing System

Heterogeneous computing systems deliver high performance and flexibility while increasing system management complexity. Resource management is particularly affected, because of the variety of computing, memory, and storage resources to handle. This paper describes Cognitive ScHeduler (CuSH), a resource management and job scheduling framework that leverages deep neural networks and reinforcement learning. To handle the complexity of heterogeneous resource scheduling, CuSH employs a two-step hierarchical solution. This solution decouples the job selection from the policy selection problem. CuSH has been evaluated using a simulator that is based on experiments executed on an IBM Power® system. Results show that CuSH outperforms traditional heuristic-based approaches on all the use cases, delivering significantly lower normalized turnaround time.