CSR: Small: A Server-Centric Approach to Data Center Networks

数据中心已成为支持在线服务的主要网络基础设施，范围从终端用户应用程序(例如，Web搜索和即时消息)到分布式系统操作(例如，GFS和MapReduceTM)。在本项目中，我们将探索一种不同的数据中心网络(DCNS)设计范例。我们采用了新的以服务器为中心的解决方案方法，而不是以交换机为中心的模式。将智能推向服务器，服务器通常具有更开放和标准化的硬件架构和软件平台。多层低成本的商用现成迷你交换机用于连接服务器。我们充分利用了ADATA中心的大量服务器。即使每台服务器只向另一台服务器多添加一个链接，我们也可以在DCN系统中获得许多链接，因为服务器数量很大。这些增加的链路不使用高端交换机，而只使用微型交换机，为增强可扩展性、服务器间容量和容错性提供了基础。采用这种方法，我们可以设计出具有高可扩展性、增强的容错性和高服务器间容量的DCN，更好地支持表现出一对一、一对多、一对所有和所有对所有通信的各种流量模式的数据中心应用。我们的设计还为通过服务器上更开放的编程平台刺激DCN软件创新打开了更多空间，同时放弃了昂贵的、专有的核心开关。在可扩展和模块化的数据中心和相关管理工具中的结果预计将表明，增加服务器的复杂性只会导致较小的系统开销，但会产生巨大的性能收益。

Data centers have become the main networking infrastructure tosupport online services, ranging from end-user applications (e.g.,Web search and IM) to distributed system operations (e.g., GFS andMapReduce). In this project, we explore a different design paradigmto data center networks (DCNs). We take a new server-centricsolution approach, in contrast to the switch-centered paradigm. Wepush intelligence into servers, which generally have more open andstandardized hardware architecture and software platforms. Multiplelayers of low-cost, commodity off-the-shelf, mini-switches are usedto connect servers. We leverage the large number of servers in adata center. Even when each server just adds one more link to therest of servers, we obtain many links in the DCN system given thelarge server population. These added links, without using high-endswitches but using only mini-switches, provide the foundation forenhancing scalability, inter-server capacity, and fault tolerance.Taking the approach, we can design DCNs with high scalability,enhanced fault-tolerance, and high inter-server capacity, and bettersupport data center applications that exhibit various trafficpatterns of one-to-one, one-to-several, one-to-all, and all-to-allcommunications. Our design also opens more space for stimulatinginnovations in the DCN software through the more open programmingplatform on servers, while abandoning expensive, proprietary coreswitches. The results in scalable and modular data centers andassociated management tools are expected to show that, the addedcomplexity into servers will incur only minor systems overhead butproduce large performance gains.