Energy Aware Task Scheduling in Data Centers

Nowadays energy consumption problem is a major issue for data centers. The energy consumption increases significantly along with its CPU frequency getting higher. With Dynamic Voltage and Frequency Scaling (DVFS) techniques, CPU could be set to a suitable working frequency during the running time according to the workload. On the other side, reducing frequency implies that more servers will be utilized to handle the given workload. It is a critical problem to make a tradeoff between the number of servers and the frequency of each server for current workload. In this paper, we investigate the task scheduling problem in a heterogeneous servers environment. To choose a suitable server among heterogeneous resources, the Benefit-driven Scheduling (BS) algorithm is designed to match the tasks to the best suitable type of server. This paper proved that the task scheduling problem based on DVFS, with the target of minimizing power consumption in a heterogeneous environment is NP-Hard. Then we proposed two heuristic algorithms based on different ideas. Power Best Fit (PBF) is based on a locally greedy manner, it always uses the least power consumption increment placement as its choice. Load Balancing (LB) uses a load balancing way to avoid over-consolidation. LB usually has a better performance than PBF, while PBF is easily turned into an online version. Compared with First Fit Decreasing (FFD) algorithm, the results show that PBF can get 12% to 13% power saving on average and LB are about 14% power saving, although PBF and LB use about 1.3 times number of servers.