SP: Very Fine-Grained Proximity Addressing

今天，连接到互联网的主机系统有两个基本标识符：IP地址和域名。这两个标识符都不能可靠地表明系统在互联网中相对于彼此的“定位”。两个IP地址或域名可能非常相似，但是它们所代表的两个系统在延迟或吞吐量方面可能相差很大。缺乏邻近“地址”是Internet架构的一个严重限制。已经有几个研究项目来解决这个限制。该领域的开创性工作是IDMaps，它提出了一种基础设施服务，可以回答关于IP地址对之间的邻近性的查询。随后提出使用全球网络定位（GNP），这是一种由叠加在互联网上的笛卡尔坐标系统形成的邻近地址。这些方法和其他方法的主要缺点是它们只能在课程粒度上操作。当延迟小于50ms时（典型的东海岸到西海岸延迟！），200%和300%的高估是很常见的。这对于网络第一人称射击游戏和紧密耦合的网格计算应用来说是完全不够的，这两种应用都需要低于50ms的延迟。如果同一校园网或isp中的参与者能够找到彼此，那么数据密集型网格计算应用程序和点对点应用程序（如文件共享）将显著减少网络负载。现有方法的另一个缺点是，它们通常需要部署专用的基础设施。虽然这对于像Akamai这样的私人公司来说是可行的，但目前还没有一个已知的公共邻近服务的商业模式。该项目将设计和实现一个系统，用于在非常细的粒度（微秒延迟和千兆吞吐量）和全球范围内计算邻近地址。这个系统必须是分布式的（点对点），要求所有节点都能作为坐标系统中的“参考点”。这是因为除非参考点在微秒内，否则不可能表示微秒延迟。然而，全局细粒度分层邻近地址的存在也带来了其他诱人的可能性。例如，这样的地址可以用作构建一种新型路由覆盖的基础，然后可以用作点对点发现机制。该项目提出了与网络路由从平面寻址到分层寻址所需的挑战相当的重要的新智力挑战。研究小组计划设计新的邻近寻址形式，这种寻址形式是分层的，而不是扁平的，不会将负载集中在几个系统上，并且足够简单，可以自动配置。研究人员还必须设计新型的P2P网络，为配置邻近地址的问题定制。这些p2p网络必须比目前的网络更简单，这样它们才能真正地按全球比例扩展，即使面对客户流失（节点加入和离开）。该项目预计将在分层邻近地址的结构及其与网络度量的关系以及如何构建非常大而简单的点对点网络方面产生广泛的成果。这个项目将产生两种形式的影响。首先，任何动态发现参与者并要求参与者在附近的分布式应用程序都将受益于非常细粒度的邻近地址。这些应用包括网络游戏和网格计算。实际上，粒度的改进可以创建新的应用程序。其次，本项目所需的点对点网络的进步可以作为其他类型的点对点应用程序的基础。具体来说，任何需要发现节点的点对点应用程序都可以发挥一定的作用（即，作为参考点），这将受益于该项目所需的规模和效率的进步。

Today host computer systems attached to the Internet have two basic identifiers: the IP address andthe Domain Name. Neither identifier reliably indicates how systems are "positioned" in the Internetwith respect to each other. Two IP addresses or Domain Names may be very similar, and yet the twosystems they represent may be far apart in terms of latency or throughput. This lack of a proximity"address" is a serious limitation of the Internet architecture.There have been several research projects to address this limitation. The pioneering work in this areawas IDMaps, which proposed an infrastructure service that could answer queries about proximitybetween pairs of IP addresses. A subsequent advancement proposed using Global NetworkPositioning (GNP)a proximity address formed from a Cartesian coordinate system super-imposedon the Internet.The primary shortcoming of these and other approaches is that they operate only at coursegranularity. At less than 50ms latency (a typical east-to-west coast latency!), over-estimates of 200%and 300% are common. This is entirely inadequate for applications such as networked first personshooter games and tightly-coupled grid computing applications, both of which require latencies wellbelow 50ms. Data intensive grid computing applications and peer-to-peer applications like filesharing would reduce network load significantly if participants in the same campus network or ISPcould find each other. Another shortcoming of existing approaches is that they typically require thatdedicated infrastructure be deployed. While this is feasible for private corporations like Akamai,there exists no known business model for a common public proximity service.This project will design and implement a system for calculating proximity addresses at very-finegranularity (microsecond latencies and gigabit throughputs) and at global scale. By necessity thissystem will be distributed (peer-to-peer), requiring that all nodes be able to serve as "referencepoints" in the coordinate system. This is because it is impossible to represent microsecond latenciesunless the reference points are within microseconds. The existence of a global fine-grainedhierarchical proximity address, however, presents other tantalizing possibilities. For instance, suchan address could be used as the basis for building a new type of routing overlay, which could then beused as a peer-to-peer discovery mechanism.This project presents significant new intellectual challenges comparable to those in network routingrequired to move from flat addressing to hierarchical addressing. The research team plans to design new forms of proximity addressing that are hierarchical rather than flat, that do not focus load on a few systems,and that are simple enough that they can be auto-configured. The researchers also must design new types of P2P networks that are customized for the problem of configuring proximity addresses. These P2Pnetworks must be simpler than current networks so that they can truly scale to global proportion evenin the face of churn (nodes joining and leaving). This project is expected to produce broadintellectual results in the structure of hierarchical proximity addresses and how they relate to networkmetrics, as well as in how to build very large and simple peer-to-peer networks.This project will have two forms of impact. First, any distributed application that dynamicallydiscovers participants, and requires that participants be nearby, will benefit from very-fine grainedproximity addresses. Such applications include networked games and Grid computing. Indeed theimprovements in granularity may enable the creation of new applications. Second, the advances inpeer-to-peer networking required for this project may serve as the basis for other types of peer-to-peer applications. Specifically, any peer-to-peer application that requires the discovery of nodes thatcan serve a certain role (i.e., act as reference points) will benefit from the advances in scale andefficiency required for this project.