A Cost-Effective Hybrid Cloud Resource Scaling Framework for Batch Processing Services
Qinzhi Zhang, Li Pan, Shijun Liu · IEEE Transactions on Network Science and Engineering · 2024
Batch processing services, like offline video processing, are pivotal in modern data analysis. Software as a Service (SaaS) providers typically purchase virtual machines (VMs) or Function as a Service (FaaS) instances, also known as function instances, from cloud providers to provision computational resources for their services. VMs offer stable performance and cost-effectiveness for continuous workloads but may incur resource waste due to idleness. Conversely, function instances, with rapid auto-scaling and fine-grained billing, excel in handling discrete workloads, albeit at a higher unit price. SaaS providers can leverage the advantages of both VMs and function instances, to achieve cost-effective service delivery while ensuring overall performance. However, due to the complexity and unpredictability of batch processing service workloads, achieving this goal is challenging. To address these issues, in this paper we propose a proximal policy optimization (PPO) based hybrid resource scaling algorithm and design a hybrid resource scaling framework. The proposed scaling framework considers the workload characteristics and performance requirements of batch processing services, adaptively making cost-optimal resource scaling decisions based on current workloads and configuration of computational resources, while ensuring the overall performance of the service. We conduct extensive simulation experiments on multiple workloads with different levels of discreteness extracted from Microsoft and Huawei datasets, and the results demonstrate that our framework can achieve optimal service cost while ensuring overall performance.