Abstract: Serverless computing is an emerging paradigm of cloud computing, which providing automatic resources provisioning and high-performance guarantees. Natures of Serverless function, like fine-grained resource allocation, short-lived lifecycles, and burst requests etc., fundamentally changed many underlying assumptions of cloud computing. Thus, the infrastructure design of performance-oriented serverless computing faces challenges. To meet SLAs requirements, Serverless platforms need to schedule tasks and allocate resources at high throughput, which presents performance constraints for platform architecture. To explore architectural implications of these natures, in this paper, we designed a novel and practical benchmarking suite, BBServerless. BBServerless dedicates to capture system-level performances impacted by the burst traffic, and helps to reveal performance bottlenecks to optimize programming and architecture designing. With a traffic generating algorithm, we conduct experiments with trace-driven simulations. Evaluation result shows that BBServerless demonstrated the correlation of system-level metrics and end-to-end performance of Serverless functions. BBServerless is implemented and publicly available at GitHub.
Authors: Yanying Lin (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China); Kejiang Ye (Chinese Academy of Sciences, China); Yongkang Li (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China); Peng Lin (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences & Shenzhen University, China); Yingfei Tang (PingAn Insurance(Group)Company of China, Ltd., China); Cheng-Zhong Xu (University of Macau, China)
Email: yy.lin1@siat.ac.cn, yekejiang@gmail.com, yk.li1@siat.ac.cn, peng.lin@siat.ac.cn, Yfei.Tang@gmail.com, czxu@um.edu.mo