DREAM: DistRibuted Energy-Aware traffic Management for Data Center Networks
DREAM: DistRibuted Energy-Aware traffic Management for Data Center Networks
复制标题
DOI：
10.1145/3307772.3328291
复制
发表时间：
2019-06
期刊：
影响因子：
--
通讯作者：
Liang Zhou;L. Bhuyan;K. K. Ramakrishnan-K.
中科院分区：
文献类型：
--
作者：
Liang Zhou;L. Bhuyan;K. K. Ramakrishnan-K.
为了在数据中心网络中节约能源，已经提出了流量整合。然而，现有的集中式流量合并方法侧重于实现最优的网络能量节约，而没有考虑响应流量变化的需要。在本文中，我们提出了一种分布式小流级流量整合框架DREAM，用于在数据中心网络中实现能量效率。Dream通过调整发送小数据流的路径选择概率，基于ECN反馈将TCP流拆分到多条路径上。这有助于选择路径，同时避免交换机上的拥塞/队列堆积。DREAM中的分布式代理在每台服务器上的Open vSwitch中实现，而无需修改数据中心网络中的应用程序、TCP或硬件交换机。使用维基百科Web服务和Facebook MapReduce流量的跟踪进行的试验台评估证明，Dream平均为数据中心网络实现了至少15.8%的节能，而最先进的方法，如Carpo和ElasticTree，分别实现了11.6%和8.4%的节能。Dream的丢包率低于0.01%，而其中最好的ElasticTree的丢包率至少为0.19%。与最先进的集中式流量整合方法相比，Dream的应用级延迟也降低了30%。
Traffic consolidation has been proposed to save energy in data center networks. However, existing centralized traffic consolidation approaches focus on achieving optimal network energy saving, without considering the need to be responsive to traffic variability. In this paper, we present DREAM, a distributed flowlet-level traffic consolidation framework for achieving energy efficiency in data center networks. DREAM splits a TCP flow across multiple paths based on ECN feedback by adapting the path selection probability for sending a flowlet. This helps to choose a path while avoid congestion/queue build-up at switches. Distributed agents in DREAM are implemented in Open vSwitch at each server without modifying applications, TCP, or the hardware switches in the data center network. Testbed evaluations using traces from Wikipedia web service and Facebook MapReduce traffic prove that DREAM on average achieves at least 15.8% energy saving for the data center network, while state-of-the-art approaches such as CARPO and ElasticTree produce 11.6% and 8.4% energy saving, respectively. The packet drop ratio in DREAM is less than 0.01% while the best among the alternatives, ElasticTree, has at least 0.19% drop ratio. DREAM also has 30% lower application-level latency than state-of-the-art centralized traffic consolidation approaches.