High-Performance Network Architecture with Speculative Scheduling for Globally Active Congestion Control

项目摘要：在本研究中，我们开发了用于多处理器、基于分布式网络的多计算机集群、存储系统和IP路由器结构的高性能互连网络的有效拥塞控制机制。将探讨允许直接涉及网络拥塞的数据包快速推进的技术。我们的方法是基于给路由器和网络接口添加最小的智能——以推测调度能力的形式——这样，当数据包被有效路由时，资源被最大限度地利用。这可以通过更有效地在拥塞区域周围引导数据包，以及更精确地检测和处理由封闭资源集上的循环相关拥塞引起的潜在死锁情况来实现。一个副产品是阻塞树/循环的全局扩展，这可能会严重降低性能。除了为该方法提供正式的理论支持外，还将通过建模和仿真进行详细的贬值。这项研究的意义是双重的。首先，通过有效地将数据包从网络的拥塞区域中分散出来，可以实现更高的持续吞吐量和更低、更可预测的延迟。也就是说，在给定的时间段内，可以通过网络发送和接收更多的数据，并且可以减少网络交付数据所需的时间，减少可变性。其次，通过有效地解决网络中的死锁，可以保证数据包的路由永远不会停止。这些重要因素使得需要在系统组件之间进行大量通信的计算机应用程序运行得更快、更可靠。

Project Abstract:In this research, we develop efficient congestion controlmechanisms for high-performance interconnection networks used inmultiprocessors, distributed network-based multicomputer clusters,storage systems, and IP router fabrics. Techniques will beexplored that allow rapid advancement of packets directly involvedin network congestion. Our approach is based on adding minimalintelligence to routers and network interfaces---in the form ofspeculative scheduling capability---such that resources are maximallyutilized while packets are efficiently routed. This can be done bymore effectively steering packets around congested areas and by moreprecisely detecting and handling potential deadlock situations arisingfrom cyclically-correlated congestion on a closed set of resources.A by-product is that global expansion of congestion trees/cycles whichcan seriously degrade performance is prevented. In addition todeveloping formal theoretical support for the approach, detailedevaluations will be carried out through modeling and simulation.The significance of this research is two-fold. First, by effectivelydispersing packets out of congested regions of the network, highersustained throughput and lower, more predictable latency can beachieved. That is, more data can be sent and received through thenetwork over a given period of time, and the time it takes for thenetwork to deliver the data can be reduced, with less variability.Second, by efficiently resolving deadlock in the network, therouting of packets is guaranteed never to come to a halt. Theseimportant factors allow computer applications which require a greatamount of communication between system components to run muchfaster and dependably.