Energy-Efficient Stable and Balanced Task Scheduling in Data Centers

It is well known that load balancing in data centers can lead to unnecessary energy usage if all servers are kept active. Usingdynamic server provisioning, the number of servers that serve requests can be reduced by turning off idle servers and thereby savingenergy. However, such a scheme, usually increases the risk of instability of server queues. In this work, we analyze the trade-offbetween energy usage and stability of servers in a data center when we balance the load by dispatching arriving jobs. We proposealgorithms to solve a stability and energy objective stochastic optimization problem with a high degree of flexibility to handle the trade-offbetween these two objectives. We consider variable size jobs to apply load balancing on selected active servers and find that the optimalsolution is an NP-hard problem. We therefore develop two computationally efficient greedy and randomized approximation schemes toachieve the trade-off between these objectives. We investigate the performance of our proposed algorithms in minimizing the risk ofqueue length growth as well as the number of active servers needed to serve jobs, and compare it with several metrics in heterogeneousload scenarios.