A GAN-based method for time-dependent cloud workload generation

To design repeatable and comparable resource management policies for data centers, researchers mainly conduct experiments in the simulation environment, which requires large-scale workload traces to simulate real scenes. However, issues related to data collection, security and privacy hinder the public availability of cloud workload datasets. Though workload generation is a promising solution, due to the unpredictable time dependency, cloud workloads are difficult to model. In light of this, we propose a novel end-to-end model for time-dependent cloud workload generation using Generative Adversarial Networks, which adopts improved Temporal Convolution Networks and Spectral Normalization to capture the time dependency and stabilize the adversarial training. Experimental results on real cloud datasets demonstrate that our model can efficiently generate realistic workloads that fulfill the diversity, fidelity and usefulness. Further, we also propose a conditional GAN which is trained with labeled data and can generate specific kind of workloads according to the input.