The energy consumption of data centers assumes a significant fraction of the world's overall energy consumption. Most data centers are statically provisioned, leading to a very low average utilization of servers. In this work, we survey uni-dimensional and high-dimensional approaches for dynamically powering up and powering down servers to reduce the energy footprint of data centers while ensuring that incoming jobs are processed in time. We implement algorithms for smoothed online convex optimization and variations thereof where, in each round, the agent receives a convex cost function. The agent seeks to balance minimizing this cost and a movement cost associated with changing decisions in-between rounds. We implement the algorithms in their most general form, inviting future research on their performance in other application areas. We evaluate the algorithms for the application of right-sizing data centers using traces from Facebook, Microsoft, Alibaba, and Los Alamos National Lab. Our experiments show that the online algorithms perform close to the dynamic offline optimum in practice and promise a significant cost reduction compared to a static provisioning of servers. We discuss how features of the data center model and trace impact the performance. Finally, we investigate the practical use of predictions to achieve further cost reductions.
翻译:数据中心的能源消耗量占世界能源消耗总量的很大一部分。 大部分数据中心都是静态提供的, 导致服务器的平均利用率非常低。 在这项工作中, 我们调查了动态上电和下电服务器的单维和高维方法, 以减少数据中心的能源足迹, 同时确保及时处理进入的工作岗位。 我们实施了平滑在线螺旋优化的算法, 以及这些算法的变异, 每轮中, 代理商都会得到一个螺旋形成本功能。 该代理商试图在每轮中尽量降低这一成本和与改变决策有关的移动成本之间取得平衡。 我们以最笼统的形式执行算法, 邀请今后研究其在其它应用领域的性能。 我们利用脸书、 微软、 阿里巴巴和 洛斯阿拉莫斯国家实验室的痕迹来评估应用正确化数据中心的算法。 我们的实验显示, 在线算法在实践上接近动态离线的最佳功能, 并承诺与静态提供服务器相比, 大幅降低成本。 我们讨论数据中心模型的特征, 并追踪其影响。 最后, 我们调查如何实际使用预测, 以进一步降低成本。