NeTS: Small: Scalable Network Virtualization for Multi-Core Systems

随着虚拟化在整个数据中心的传播，数据中心内的通信端点正在变成虚拟机(VM)，而不是物理服务器。因此，用于数据包处理和安全的网络设施必须一直运行到VM。物理服务器内的虚拟机之间的这种最后一跳交换已成为数据中心网络的一个至关重要的组件。本项目将探索服务器/交换机集成的设计空间，其中基于软件和硬件的交换可以更有效地直接集成到服务器中。更具体地说，它将分离交换操作，并将其实施分布在整个虚拟系统中。该项目将由以下主要部分组成：-通过使用基于流的方法消除这些管理费用并实现软件中的高效交换的分离的软件交换架构实现高效的软件交换-明智地选择和使用硬件加速。-通过让网络控制平面了解硬件拓扑、监视网络流量、指示管理程序将通信的VM彼此靠近，机架成为服务器。随着对数据中心容量的需求继续以令人难以置信的速度增长，将物理机数量降至最低变得至关重要。这项研究将改变在未来系统上实施联网的方式，使每台物理机能够有效使用100台虚拟机。物理机器的这种减少可能会产生重大的社会和经济影响。如果美国数据中心使用虚拟化技术实现英特尔预测的电力削减的一半，我们将节省35亿美元的电力，并防止每年释放3.22亿磅二氧化碳。

As virtualization spreads throughout the data center, the communication endpoints within the data center are becoming virtual machines (VMs), not physical servers. Therefore, the network facilities for packet processing and security must operate all the way to the VM. This last hop switching among virtual machines within the physical server has become a critically important component of the data center network. This project will explore the design space of server/switch integration, in which software and hardware based switching are more efficiently integrated directly into the server. More specifically, it will decouple the switching operations and distribute their implementation throughout the virtualized system. This project will be comprised of the following thrusts: - Efficient Software Switching via a decoupled software switching architecture using a flow-based approach that eliminates these overheads and that enables efficient switching in software - Judicious choice and use of hardware acceleration. - The Rack Becomes the Server via having the network control plane learning the hardware topology, monitoring the network traffic, instructing the hypervisor to move communicating VMs closer to each other. As the demand for data center capacity continues to increase at an incredible rate, it is becoming critical to minimize the number of physical machines. This research will transform the way in which networking is implemented on future systems, enabling the effective use of 100s of virtual machines per physical machine. This reduction in physical machines can have significant societal and economic impacts. If U.S. data centers used virtualization technology to achieve just half of the power reductions predicted by Intel, we would save $3.5 billion of electricity and prevent the release of 322 million pounds of CO2 per year.