Long-Term Multi-objective Task Scheduling with Diff-Serv in Hybrid Clouds

With the speedy development of E-commerce, requests over the internet from intensive users are soaring, especially in global online shopping festivals. In order to meet the increasing demands of temporary capacity and reduce daily expenses, hybrid clouds are often used, and the task scheduling problem with multi-objectives is further investigated. In this paper, we firstly build a differentiated-service (Diff-Serv) task scheduling model, and formulate a dynamic programming problem, where the state space is too large to be solved by exhaustive iterations. Therefore, we carefully design the value approximation function, and with reference to the reinforcement learning theory, we put forward an approximate dynamic programming (ADP) algorithm so as to conduct the long-term optimization for performance benefit, energy and rental costs. Furthermore, both scheduling quality and scheduling speed are taken into consideration in this algorithm. Experiments with both random synthetic workloads and Google cloud trace-logs are conducted to evaluate the proposed algorithm, and results demonstrate that our algorithm is effective and efficient, especially under bursty requests.