CAREER: High Performance Packet Switches

Internet同时面临着两个问题：需要更快的交换/路由能力来承载持续增长的用户数据，以及需要为实时流量引入服务质量（QoS）。 每个问题都可以独立解决-最快的单级交换机和路由器可以使用输入排队交叉开关，而不是受内存带宽限制的共享内存系统。 并且可以使用许多公知的技术来提供QoS，诸如加权公平协商（WFQ）。 然而，到目前为止，这两种解决方案一直是相互排斥的;所有关于提供QoS的算法的工作都要求交换机和路由器使用共享存储器或输出路由器。下一代互联网（NGI）将要求极大地增加容量，并保证卓越的服务质量。 基于以前的高性能交换机和路由器的研究，建议的工作将遵循三个途径，试图解决同时增加网络容量和提供QoS的问题：1。交换机加速。 使用一种名为“通过加速进行精确仿真”的新技术，可以构建交叉开关，精确仿真输出队列和集中式共享内存开关，但使用的内存运行速度仅为外部线路的两倍。 这种有前途的技术使交换机和路由器能够以比以前高得多的速度提供有保证的QoS。 在这项工作中，新的实用算法将被开发，研究和原型，目标是使这项技术可用于整个行业，2。Fork-Join路由器 这是一种新的高度并行的单级IP路由器的架构，总容量超过1 Tb/s。 同时使用多种形式的并行性，系统的不同部分可以以不同的速度操作，从而允许特定实现最佳地利用可用的交换和转发技术。 该路由器具有高度的容错性，当子系统发生故障时，它会优雅地降级。 也许最重要的是，fork-join路由器可以为提供精确QoS的T比特路由器使用“加速”技术。流处理器。 网络交换机和路由器需要快速处理元件，以便在数据“流过”时对其进行处理。“传统的重流水线加载存储RISC CPU不能很好地处理流数据。 需要一种针对流数据优化的处理器架构，其中CPU是更大系统流水线的一部分，并且其中可以非常高的速度处理具有空间局部性而不是时间局部性的数据。 这样的设备将使路由器中的转发决策更快，执行更快的调度算法以提供QoS，并执行网络数据的快速转码。 因此，一个目标，拟议的研究是转让技术的建设者的“电信级”交换和路由设备，并鼓励迅速采用。

The Internet is facing two problems simultaneously: there is a need for faster switching/routingcapacity to carry the continued growth of user data, and there is a need to introduce guaranteedqualities of service (QoS) for real-time traffic. Each problem can be solved independently - the fastest single-stage switches and routers can be made using input-queued crossbar switches, instead of shared memory systems limited by memory bandwidth. And QoS can be provided using numerous well-known techniques, such as weighted fair queueing (WFQ). However, until now, these two solutions have been mutually exclusive; all of the work on algorithms that provide QoS has required that switches and routers use shared memory or output queueing.The Next Generation Internet (NGI) will require hugely increased capacity, and distinguished and guaranteed qualities of service. Building on previous research on high performance switches and routers, the proposed work will follow three avenues in an attempt to solve the problem of simultaneously increasing network capacity and providing QoS:1. Switch Speedup. Using a new technique called "precise emulation via speedup", cross-bar switches can be built that precisely emulate output queued and centralized shared memory switches, yet use memories that run at only twice the speed of the external line. This promising technique enables switches and routers to provide guaranteed QoS at speeds much higher than before. In this work, new practical algorithms will be developed, studied and prototyped with the goal of making this technique available to the industry at large,2. The Fork-Join Router. This is a new highly parallel architecture for single-stage IP routers with aggregate capacities in excess of 1Tb/s. Using multiple forms of parallelism simultaneously, different parts of the system may operate at different speeds, allowing a particular implementation to make best use of the available switching, and forwarding technology. The router is highly fault-tolerant, degrading gracefully when sub-systems fail. Perhaps most importantly, the fork-join router can use the "speedup" technique for Terabit routers that provide precise QoS guarantees.3. The Streaming Processor. Network switches and routers require fast processing elements that process data as it "streams through." Streaming data is not processed well by traditional heavily-pipelined load-store RISC CPUs. There is a need for a processor architecture optimized for streaming data, in which the CPU is part of a larger system pipeline, and in which data with spatial, but not temporal, locality can be processed at very high speed. Such a device would enable faster forwarding decisions in routers, performfaster scheduling algorithms to provide QoS, and perform fast transcoding of net-work data.An important goal is to see the research used by industry and in the NGI. Therefore, a goal ofthe proposed research is to transfer the technology to the builders of "carrier-class" switching and routing equipment, and to encourage rapid adoption.