Energy-efficient content delivery networks using cluster shutdown
Energy-efficient content delivery networks using cluster shutdown
复制标题
DOI：
10.1109/igcc.2013.6604510
复制
发表时间：
2013-06
期刊：
影响因子：
--
通讯作者：
Vimal Mathew;R. Sitaraman;Prashant J. Shenoy
中科院分区：
文献类型：
--
作者：
Vimal Mathew;R. Sitaraman;Prashant J. Shenoy
内容分发网络（CDN）是一类重要的互联网规模分布式系统，它将网页、流媒体和应用内容分发给终端用户。一个商业CDN可能包含部署在全球数千个集群中的数十万台服务器，并且在为其服务器供电和冷却方面产生大量能源成本。由于能源成本是CDN总运营费用的重要组成部分，我们提出并探索一种名为集群关闭的新技术，该技术关闭部署在数据中心内的CDN的整个服务器集群。通过这样做，集群关闭不仅节省了服务器消耗的电力，还节省了冷却这些服务器所需的电力。我们提出了一种基于服务器和冷却设备的实际功率模型的集群关闭算法，该算法可作为CDN全局负载均衡器的一部分来实现。我们使用来自一个大型商业CDN的大量真实世界跟踪数据来评估我们的技术，结果表明集群关闭可将系统范围内的能源使用量减少67%。此外，在不牺牲带宽成本或终端用户性能的情况下，大部分能源节省是可以实现的。此外，即使每个集群每天最多限制关闭一次，也能实现79%的最优节省，从而减少所需的运营开销。最后，我们认为在CDN环境中，集群关闭相对于经过充分研究的服务器关闭技术具有内在的架构优势，并表明在广泛的运行范围内它比服务器关闭节省更多能源。
Content delivery networks (CDNs) are an important class of Internet-scale distributed systems that deliver web, streaming, and application content to end users. A commercial CDN could comprise hundreds of thousands of servers deployed in over thousand clusters across the globe and incurs significant energy costs for powering and cooling their servers. Since energy costs are a significant component of the total operating expense of a CDN, we propose and explore a novel technique called cluster shutdown that turns off an entire cluster of servers of a CDN that is deployed within a data center. By doing so, cluster shutdown saves not just the power consumed by the servers but also the power needed for cooling those servers. We present an algorithm for cluster shutdown that is based on realistic power models for servers and cooling equipment and can be implemented as a part of the global load balancer of a CDN. We evaluate our technique using extensive real-world traces from a large commercial CDN to show that cluster shutdown can reduce the system-wide energy usage by 67%. Further, much of the energy savings are obtainable without sacrificing either bandwidth costs or end-user performance. In addition, 79% of the optimal savings are attainable even if each cluster is limited to at most one shutdown per day, reducing the required operational overhead. Finally, we argue that cluster shutdown has intrinsic architectural advantages over the well-studied server shutdown techniques in the CDN context, and show that it saves more energy than server shutdown in a wide range of operating regimes.