Learning to Schedule Multi-Server Jobs with Fluctuated Processing Speeds

Abstract

Multi-server jobs are imperative in modern cloud computing systems. A multi-server job has multiple components and requests multiple servers for being served. How to allocate restricted computing devices to jobs is a topic of great concern, which leads to the job scheduling and load balancing algorithms thriving. However, current job dispatching algorithms require the service rates to be changeless and knowable, which is difficult to realize in production systems. Besides, for multi-server jobs, the dispatching decision for each job component follows the All-or-Nothing property under service locality constraints and resource capacity limits, which is not well supported by mainstream algorithms. In this paper, we propose a dispatching algorithm for multi-server jobs that learns the unknown service rates and simultaneously maximizes the expected Accumulative Social Welfare (Asw). We formulate the Asw as the sum of utilities of jobs and servers achieved over each time slot. The utility of a job is proportional to the valuation for being served, which is mainly impacted by the fluctuating but unknown service rates. We maximize the Asw without knowing the exact valuations, but approximate them with exploration-exploitation. From this, we bring in several evolving statistics and maximize the statistical Asw with dynamic programming. The proposed algorithm is proved to have a polynomial complexity and a State-of-the-Art regret. We validate it with extensive simulations and the results show that the proposed algorithm outperforms several benchmark policies with improvements by up to 73%, 36%, and 28%, respectively.