In a heterogeneous workflow environment, the uncertainty of task execution times, dynamic resource changes, and task dependencies’ evolution pose significant scheduling challenges. This paper investigates how to make intelligent and adaptive scheduling decisions in a constantly changing heterogeneous cloud environment. We propose a novel scheduling approach, Transformer-enhanced DQN (T-DQN) that combines the strengths of reinforcement learning and the Transformer model into a hybrid strategy. This method leverages the ability of reinforcement learning to handle uncertainty and dynamics in the decision-making process while integrating the advantages of the Transformer model in dealing with long sequences and complex relationships. Our experimental evaluation shows that the T-DQN algorithm outperforms existing algorithms consistently in real-world workflow, dynamic scenarios, and high-load environments. T-DQN reduces makespan by up to 13.66%, improves energy by about 16.65%, and improves cost by 44.72% compared to the six other approaches. This performance is particularly significant in high-load environments, where T-DQN’s adaptability and scalability minimize failure rates and optimize resource management, affirming its suitability as a robust solution to complex cloud computing challenges.

The spatial distribution and long-time variation of the deep-developed boundary layer are not well understood in arid and semi-arid regions of northwest China. ERA-Interim (ECMWF Reanalysis data, ECMWF: European Centre for Medium-Range Weather Forecasts) were used to study the deep-developed boundary layer in the five representative areas in summer and then the Weather Research Forecast (WRF) model was applied to simulate and verify its applicability. The results show that the boundary layer heights in the five representative areas are higher in late spring and summer (the highest is 2485~3502 m in June) and lower in autumn, winter and early spring (the lowest is 758~907 m in December). The seasonal variations of the boundary layer height are smaller at 02:00 BJT and 08:00 BJT, while the variations are relatively larger at 14:00 BJT and 20:00 BJT. The atmospheric boundary layer, with heights over 4000 m, generally exists in late spring and summer. The boundary layer heights are higher in the arid region than in the semi-arid region and the deep-developed boundary layer lasts longer in the arid region. The boundary layer heights present reductions from the northwest to the southeast, except for Minqin in the middle north. The numerical simulation results show that there is a significant difference between different combinations of parameterization schemes to simulate the deep-developed boundary layer in these areas. The combination Goddard+SLAB+ACM2 performs better in the extreme arid area, Dunhuang, and the arid areas, Jiuquan and Minqin, whereas the simulation effect of the combination Dudhia+Noah+ACM2 is better in the semi-arid areas, Yuzhong and Lanzhou. The difference between the schemes is related to the determination of the boundary layer height.

Scheduling workflows with complex dependencies under dynamic resource availability in heterogeneous clouds remains highly challenging. Static and heuristic methods often fail to adapt to workload and resource changes, resulting in higher makespan, lower utilization, and increased cost. We propose GDST-PPO, a graph-based dynamic scheduling framework that encodes workflow DAGs with a GNN, models temporal and resource contexts via a Transformer with sparse attention, and optimizes decisions using PPO with a multi-objective reward. We perform extensive experiments on five benchmark workflows under realistic heterogeneous cloud settings. Across six baselines, GDST-PPO achieves up to 10.3% lower makespan, 18.1% higher resource utilization, and an 11.4% gain in overall score. In absolute terms, it attains a makespan of 7,168,820s, resource utilization of 39.55%, and an overall score of 6.62 on our benchmark, demonstrating efficient, flexible, and cost-effective cloud workflow management.

Hybrid cloud deployments combine the low-latency communication of on-premises high-performance computing (HPC) clusters with the elasticity of public clouds, yet current numerical weather prediction (NWP) runtimes typically exploit only one of these capabilities. MPI lacks dynamic scalability, while Spark incurs significant serialization overhead. To address this limitation, we propose a zero-copy hybrid runtime that connects the persistent MPI COMM_WORLD with Apache Spark via Arrow Flight, eliminating JVM–native data copies and reducing cross-layer transfer latency. We further develop a fuzzy-logic scheduler fuzzy-based hybrid cloud task scheduling (FbTS) coupled with a LightGBM runtime predictor to jointly optimize turnaround time, energy consumption, and cloud cost under a VM-hour pricing model. Experiments with 100 weather research and forecasting (WRF) forecasting runs and workloads up to 100 GB show that the proposed system reduces wall-clock time by 83.7% when scaling from 16 to 128 cores while maintaining 0.91 parallel efficiency. On a 1000-node hybrid cloud, FbTS achieves a 22% runtime reduction and 17% energy savings compared with the best competing heuristic.