•  
  •  
 

Abstract

Cloud computing continues to consume increasing amounts of energy, adding to operating costs and making sustainability and service reliability harder to maintain. Resource management, however, is still largely reactive in many systems: decisions are made after workload or renewable-energy conditions have already changed, with little use of what is likely to happen next. GreenWaveTransformer is designed around this gap. It combines a wavelet-based Transformer with an offline reinforcement-learning (RL) controller so that resource-management decisions can be informed by expected operating conditions rather than by the current state alone. A discrete wavelet transform is used to represent nonlinear and nonstationary workload dynamics, while structured pruning and quantization reduce Transformer inference overhead. The RL agent is trained entirely from historical data, avoiding exploration on a live system. Evaluation was conducted in simulation using Google and Alibaba cluster traces, with synthetic renewable-energy inputs cross-validated against measured telemetry from the NREL National Solar Radiation Database. Compared with forecasting-only and RL baselines, GreenWaveTransformer reduced forecasting-module inference energy by 28%, total simulated energy consumption by 22.3%, and SLA violations by 31%. Boundedness and nominal recovery held across all simulated stress scenarios, though deployment claims require validation on physical infrastructure.

Keywords

Data center sustainability, Discrete wavelet transforms, Offline reinforcement learning, Service-level agreement (SLA), Transformer models

Subject Area

Computer Science

Article Type

Article

First Page

3393

Last Page

3411

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Download

Share

 
COinS