Conditional Generative Adversarial Network Based Workload Generation for Cloud Cluster

Improving cluster utilization by scheduling and decision-making is a long-standing research problem for cloud vendors. The workload models can improve decision-making by providing a provision of the future workload for the public cloud scheduler. However, capturing the correlations in real traces was proven to be hard in former research. In this paper, we introduce a conditional generative adversarial network (CGAN) based model, which can provide a long-time provision for the cloud cluster. In the proposed approach, the workload model is generated by three-stage training with conditional GAN, which is not limited by the assumption of Poisson distribution. Besides, the conditional representation is redesigned based on Time2Vec, which makes the inter-job correlations among virtual machines (VMs) can be modeled accurately. According to our validation, the job arrival model accuracy is raised from 52.7% to 80.3% compared with the Poisson regression method, which indicated that the proposed CGAN-based model is a more universal and accurate generative model for large-scale cloud clusters.