Rolling Horizon Scheduling for Scientific Cloud Workloads Via MDP and LSTM
WANG Ying
WANG Bo
JIANG Gaoxiu
ZHU Li
Abstract：The explosive growth of computeintensive scientific workloads—such as AI training and bigdata analytics—poses triple challenges to cloud schedulers:highly dynamic task arrivals,pronounced resource heterogeneity and conflicting optimization objectives.To overcome the limited responsiveness and objective entanglement of existing approaches,this paper presents PRHS-MDP,a predictive rolling horizon scheduling algorithm formulated as a Markov Decision Process(MDP).An LSTM network is employed to forecast multistep node loads and augment the system state with foresight.A rollingwindow optimizer then maximizes a composite reward that jointly considers completion rate,mean response time,and energy consumption.Extensive simulations on a heterogeneous CloudSim testbed demonstrate that PRHS-MDP outperforms FIFO,MinMin,HEFT and PSO:task completion rate increases by 18%,mean response time drops by 23%,and total energy consumption is reduced by 11%,while preserving fast con-vergence and high robustness.The proposed method offers a practical and theoretically sound path towards intelligent,selfadaptive scheduling for scientific cloud platforms.
Keywords：cloud computingtask schedulingMarkov decision processload prediction
Publication Date：2025-11-20
Online Publishing Date：2026-01-28(First online date of this platform, not the publication date of the document)
Pages：5( 128-132 )