UDL: a cloud task scheduling framework based on multiple deep neural networks

Abstract Cloud task scheduling and resource allocation (TSRA) constitute a core issue in cloud computing. Batch submission is common user deployment mode computing systems. In this mode, it has been challenge for systems to balance the quality of service revenue provider (CSP). To end, with multi-objective optimization (MOO) minimizing latency energy consumption, we propose TSRA framework based on deep learning (DL). The system solves problems multiple queues virtual machine (VM) clusters by uniting neural networks (DNNs) as scheduler system. DNNs are divided into exploration part exploitation part. At each time step, model saves best outputs all policies from DNN experienced sample memory pool (SMP), periodically selects random training samples SMP train We designed united (UDL) algorithm framework. Experimental results show that UDL can effectively solve MOO problem tasks, performs better than benchmark algorithms such heterogeneous distributed (HDDL) terms performance.