Self-Adaptive Multi-Cloud Job Scheduling Using Deep Reinforcement Learning and Workload Forecasting

Abstract

The elastic, heterogeneous, and price-volatile nature of modern multi-cloud ecosystems makes job scheduling a persistently difficult problem. Static heuristics and metaheuristics optimize against a single snapshot of the system and degrade sharply when demand spikes, spot-market prices shift, or virtual machine (VM) availability changes across providers. This paper proposes SAMC-DRL, a self-adaptive multi-cloud scheduling framework that couples a deep reinforcement learning (DRL) scheduler with a short-horizon workload-forecasting module. The forecasting module is an attention-augmented bidirectional Long Short-Term Memory (Attention-BiLSTM) network that predicts aggregate resource demand from historical traces; its predictions are folded into the scheduler’s state representation so that placement decisions are made proactively rather than reactively. The scheduler is trained with Proximal Policy Optimization (PPO) using an adaptive, multi-objective reward that jointly balances makespan, monetary cost, energy consumption, and service-level-agreement (SLA) violations, with weights that self-tune to the observed operating regime. We evaluate SAMC-DRL on two public production traces—the Google Borg cluster trace and the Alibaba cluster trace—within an extended CloudSim simulation spanning three emulated cloud providers. Across workloads of 100–4000 concurrent tasks, SAMC-DRL reduces makespan by up to 11.3%, energy by up to 12.5%, and monetary cost by up to 14.1% relative to PPO without forecasting and lowers the SLA-violation rate from 8.4% to 3.2% relative to a Deep Q-Network (DQN) baseline. The Attention-BiLSTM forecaster attains a mean absolute percentage error (MAPE) of 5.8%, outperforming ARIMA, support vector regression (SVR), LSTM, GRU, and a vanilla Transformer. An ablation study confirms that workload forecasting and the adaptive reward are the two largest contributors to the observed gains. The results indicate that forecasting-aware, self-adaptive DRL is a practical path toward cost- and energy-efficient scheduling in operational multi-cloud platforms. All scheduling results are averaged over ten independent seeds and are statistically significant under Holm-corrected paired t-tests and exact Wilcoxon signed-rank tests, and additional experiments characterize the framework under flash-crowd, aperiodic and adversarial anti-forecast bursts, under 30-day forecast drift, and across DAG workloads of differing depth and branching factor.