通用拥塞控制及其在Linux内核中的实现

目前Internet上的拥塞控制采用以TCP为核心的基于窗口技术的端到端（end-to-end)控制方法,具有丢包、响应速度慢等缺陷,造成网络资源的极大浪费。文中提出的通用拥塞控制方法通过对拥塞的早期检测和在相邻节点间交换拥塞控制信息等方法,使网络能对拥塞做出快速响应,从而有效地克服了端到端拥塞控制方法的缺陷,实验结果表明,采用通用拥塞控制的网络在吞吐量和抖动等性能上均优于端到端拥塞控制。     

全局主动带宽调整算法

近年来，UDP等非响应流抢占TCP等响应流的现象越来越严重。提出全局主动带宽调整算法（GABA），通过网络中核心路由器与边缘路由器之间的协作，在核心路由器获取非响应流的信息，在边缘路由器阻止非响应流的大量进入。通过NS2仿真表明，GABA可以有效提高网络利用率和带宽分配的公平性。     

Interleaved Traffic Splitting: A promising technique to solve False Timeout

In this paper, we analyze the False Timeout (FTO) problem that TCP flows suffered in OBS networks and propose Interleaved Traffic Splitting (ITS) to solve the problem. We show that the collision loss of ACK bursts may also cause FTO, which has been neglected by previous researchers. As TCP Sack is designed to recover from multiple packet losses, we find that a Sack flow can recover its data transmission without triggering retransmission timeout if packet loss is not too severe. This suggests limiting the number of packets carried by a single burst. In ITS, this is achieved by carrying out a parallel, interleaved burst assembly process at each ingress OBS edge router. We show that ITS tackles the FTO problem caused by the loss of both data and ACK bursts. Unlike other existing approaches, the ITS scheme requires only modification to the assembly/de-assembly mechanisms implemented at the edge routers. It does not add complexities to the network management or the router design. In this paper, both analytical and simulation results are presented to demonstrate the performance improvement of ITS. Our analysis further shows that the use of two parallel bursts is optimum for ITS.     

A novel TCP with dynamic Burst-Contention Loss notification over OBS networks

In this paper, a novel congestion control scheme with dynamic Burst-Contention Loss notifications in Optical Burst Switching (OBS) networks is proposed. The proposed scheme, called TCP-BCL, aims to handle various OBS bursty conditions that negatively affect TCP throughput performance and fairness. The basic design principle of the scheme is to tune the congestion control parameters α and β such that the congestion window sizes in the corresponding TCP senders can be adjusted with an explicit notification from the OBS edge node. The performance impact on TCP in terms of burst dropping due to random contention, which is also known as false congestion detection is considered and investigated. An analytical model is developed and further verified through extensive simulation.     

OBS中TCP业务的分段指示拥塞控制策略

传统的TCP拥塞控制算法主要是针对电通信网络中包交换机制提出的,当这种拥塞控制算法应用到光突发交换（OBS）中会出现很多问题,甚至会使网络性能急剧恶化。为了改善传统的TCP拥塞控制算法在OBS网络中的性能,提出了一种分段指示拥塞控制技术,它能根据光网络链路的占用情况,在OBS边缘节点随机地标记不同TCP流的数据包以阻止网络拥塞。该方法不仅能对网络拥塞迅速地做出反应,而且还能应对OBS的假超时现象（FTO）,进一步改善OBS网络的性能。     

区分服务网络基于覆盖的拥塞管理方案

提出一种面向区分服务网络确保转发的覆盖式拥塞管理方案,基本思想是利用控制分组在网络的入口和出口节点之间传递网络服务质量的状态信息,入口节点利用加性增加和显性降低的算法调节聚集通信量的发送速率.实验结果表明,与标准的区分服务网络相比,该方案能在聚集之间公平地分配带宽并能显著地降低分组丢失率.     

一种核心无状态保存的自适应成比例公平带宽分配机制

提出了一种核心无状态的自适应比例公平带宽分配机制CSPAFA（core stateless proporitonal adaptive fair allocation)，在边界路由器完成基于每个流的状态处理，将所有流分成标记流和非标记流两种业务类型，采用DPS（dynamic packet state)技术将有关信息编码进IP分组头，在核心将输出链路带宽分成两部分，核心根据当前的网络负荷对标记流按服务规格成比例的分配输出链路带宽，对未标记流公平分配带宽，并且能自适应地调整两类业务的带宽共享比例，最后，给出了在NS网络仿真环境下的仿真实验结果。     

确保服务中一种基于动态阈值的数据包标记算法

确保转发过程中带宽享用的公平性问题一直是区分服务网络研究的热点，影响这种公平性的因素包括回路响应时间RTT、数据包大小、目标速率及聚流中包含的单流数量等．确保服务的实现依赖于在边界路由器执行的数据包标记策略和在核心路由器执行的队列管理策略，基于动态阈值的数据包标记策略DTBM的目的就是处理异质的TCP流之间带宽享用的公平性问题。DTBM通过测量局部吞吐量来调整标记算法中的阈值，以改变不同颜色的标记概率从而达到公平带宽享用的目的。DTBM的主要优点在于其实现简单、对参数不是很敏感并且对端结点主机是“透明”的，仿真实验表明，和其他几种标记算法相比，DTBM能有效地消除上述因素的影响，具有更好的公平性。     

Can high-speed networks survive with DropTail queues management?

It has been observed that TCP connections that go through multiple congested links (MCL) have a smaller transmission rate than the other connections. Such TCP behavior is a result of two components (i) the cumulative packet losses that a flow experiences at each router along its path; (ii) the longer round trip times (RTTs) suffered by such flows due to non-negligible queueing delays at congested routers. This double “bias” against connections with MCLs has been shown to approximate the so-called minimum potential delay fairness principle in the current Internet. Despite the recent proliferation of new congestion control proposals for TCP in high-speed networks, it is still unclear what kind of fairness principle could be achieved with such newly proposed congestion control protocols in high-speed networks with large-delays. Studies already show that some high-speed TCP variants may cause surprisingly severe RTT unfairness in high-speed networks with DropTail routers.This paper studies the problem of unfairness in high-speed networks with some well-known high-speed TCP variants in presence of multiple congested links and highlights the severity of such unfairness when DropTail queue management is adopted.Through a simple synchronized loss model analysis, we show how synchronized losses with DropTail in high-speed networks could lead to severe RTT unfairness and drop probability (DP) unfairness; while random marking AQM schemes, which break the packet loss synchrony mitigate such unfairness dramatically by ensuring that the packet loss probability of a flow is the sum of the loss probabilities on the congested routers it crosses.Extensive simulations are carried out and the results support our findings.     

Provision of differentiated performance in Optical Burst Switching networks based on burst assembly processes

Optical Burst Switching (OBS) has been designed as a practical solution for the high-speed transport network using enabling optical technologies, which can work without optical buffering or pure optical signal processing in the intermediate nodes. As a collection of many packets, optical bursts exhibit different traffic characteristics in the bufferless optical core networks compared to the packet-level traffic in the conventional electronic switching systems. However, many OBS studies often neglect this issue by assuming the inputs to be optical bursts when analyzing the system performance. This paper demonstrates that the loss performance of optical bursts is dramatically different when considering burst assembly process at the edge nodes and the connectivity of the core nodes. We develop some analytical models which can model the loss performance more accurately than the traditional methods. Based on our observations, we propose a scheme to support differentiated loss performances for optical bursts by modifying the burst assembly parameters at the source edge nodes. Our scheme is implemented at the edge nodes and there is no specific requirement to the core nodes. Thus, our design provides good flexibility and scalability in the heterogeneous Internet.     

Differentiated Bandwidth Allocation with TCP Protection in Core Routers

Differentiated Services (DiffServ) networks categorize routers into edge routers and core routers. In core routers, one of the technological challenges is how to implement differentiated bandwidth allocation and TCP protection together with low complexity. We present an Active Queue Management (AQM) scheme called CHOKeW. A method is borrowed from a previous scheme, CHOKe, which draws a packet at random from the buffer, compares it with the arriving packet, and drops both if they are from the same flow. CHOKeW enhances the drawing function by adjusting the maximum number of draws based on the priority of the new arrival and the current status of network congestion. With respect to the number of flows, both the memory-requirement complexity and the per-packet-processing complexity for CHOKeW is O(1). An analytical model and multiple simulations are used to explain and evaluate CHOKeW. We show that CHOKeW is able to 1) support differentiated bandwidth allocation; 2) provide the flows in the same priority with better fairness than other conventional stateless AQM schemes such as RED and BLUE; 3) maintain high link utilization as well as short queue length; and 4) protect TCP flows by restricting the bandwidth share of high-speed unresponsive flows.     

Monitoring Edge-to-Edge Traffic Aggregates in Differentiated Services Networks

Differentiated Services (DiffServ), which are currently being standardized in the IETF DiffServ working group, is a solution that can provide different qualities of service to different network users. DiffServ aggregates network packets at edge routers and forwards the aggregated packets to core routers with different priorities. In this paper, we propose methods using the SNMP framework for monitoring edge-to-edge traffic aggregates in a DiffServ domain, which consists of a set of DiffServ-enabled routers. In order to manage each DiffServ router, we have analyzed the DiffServ MIB and instrumented it in the router. Further, we propose monitoring behaviors of edge-to-edge traffic aggregates by combining topology and performance information from MIB II and DiffServ MIB. Construction procedures and graphical representation of the edge-to-edge traffic aggregates are explained in detail. We also extend our efforts to implement a DiffServ domain monitoring system that monitors a set of DiffServ-enabled routers and traffic aggregates between every edge router pair. We believe that the proposed monitoring methods can serve as useful building blocks for managing DiffServ networks.     

Minimizing the system impact of router faults by means of reconfiguration and adaptive routing

To tolerate faults in Networks-on-Chip (NoC), routers are often disconnected from the NoC, which affects the system integrity. This is because cores connected to the disabled routers cannot be accessed from the network, resulting in loss of function and performance. We propose E-Rescuer, a technique offering a reconfigurable router architecture and a fault-tolerant routing algorithm. By taking advantage of bypassing channels, the reconfigurable router architecture maintains the connection between the cores and the network regardless of the router status. The routing algorithm allows the core to access the network when the local router is disabled.Our analysis and experiments show that the proposed technique provides 100% packet delivery in 100%, 92.56%, and 83.25% of patterns when 1, 2 and 3 routers are faulty, respectively. Moreover, the throughput increases up to 80%, 46% and 33% in comparison with FTLR, HiPFaR, and CoreRescuer, respectively.     

Performance analysis of timer-based burst assembly with slotted scheduling for optical burst switching networks

In this paper, we analyze the performance of a timer-based burst assembly for optical burst switching (OBS) networks. In our analytical model, an ingress edge node has multiple buffers where IP packets are stored depending on their egress edge nodes, and bursts are assembled at the buffers in round-robin manner. Moreover, bursts are transmitted in accordance with slotted scheduling where each burst transmission starts at the slot boundary. We construct a loss model with two independent arrival streams, and explicitly derive the burst loss probability, burst throughput, and data throughput. In numerical examples, we show the effectiveness of our analysis in comparison with the Erlang loss system. It is shown that our model is quite useful for an OBS network with a large number of input and output links.     

Novel reinforcement learning-based approaches to reduce loss probability in buffer-less OBS networks

Optical Burst Switching (OBS) is a promising switching paradigm for the next generation Internet. A buffer-less OBS network can be implemented simply and cost-effectively without the need for either wavelength converters or optical buffers which are, currently, neither cost-effective nor technologically mature. However, this type of OBS networks suffers from relatively high loss probability caused by wavelength contentions at core nodes. This could prevent or, at least, delay the adoption of OBS networks as a solution for the next generation optical Internet. To enhance the performance of buffer-less OBS networks, we propose three approaches: (a) a reactive approach, called Reinforcement Learning-Based Deflection Routing Scheme (RLDRS) that aims to resolve wavelength contentions, after they occur, using deflection routing; (b) a proactive multi-path approach, called Reinforcement Learning-Based Alternative Routing (RLAR), that aims to reduce wavelength contentions; and (c) an approach, called Integrated Reinforcement Learning-based Routing and Contention Resolution (IRLRCR), that combines RLAR and RLDRS to conjointly deal with wavelength contentions proactively and reactively. Simulation results show that both RLAR and RLDRS reduce, effectively, loss probability in buffer-less OBS networks and outperform the existing multi-path and deflection routing approaches, respectively. Moreover, simulation results show that a substantial performance improvement, in terms of loss probability, is obtained using IRLRCR.     

区分服务网络基于测量的接纳控制方案的设计与应用

提出了一种分布式可扩展的接纳控制方案，其目的是为区分服务网络提供端到端服务质量（Quality of Services,QoS)保证，该方案主要由以下部分构成：（1）连接接纳控制协议，主要负责主机和网络节点以及网络节点和带宽代理之间的信息传送，实现对连接请求的串行操作；（2）位于网络核心节点的可用带宽估计算法；（3）位于网络边缘节点的接纳控制算法；给出了该方案在视频传输方面的应用实例；利用多种网络拓扑结构和QoS指标评价该方案的性能，实验结果表明该方案能准确地控制可接纳区域和提高网络资源的利用率。     

Quantitative adaptive RED in differentiated service networks

This paper derives a quantitative model between RED (Random Early Detection ) maxp and committed traffic rate for token-based marking schemes in DiffServ IP networks.Then, a DiffServ Quantitative RED( DQRED) is presented ,which can adapt its dropping probability to marking probability of the edge router to reflect not only the sharing bandwidth but also the requirement of performance of these services,Hence,DQRED can cooperate with marking schemes to guarantee fairness between different DiffServ AF class services,A new marking probability metering algorithm is also proposed to cooperate with DQRED ,Simulation results verify that DQRED mechanism can not only control congestion of DiffServ network very well,but also satisfy different quality requirements of AF class service.The performance of DQRED is better than that of WRED.     

On burst rescheduling in OBS networks with partial wavelength conversion capability

The core nodes in an optical burst switching (OBS) network are normally equipped with wavelength converters (WCs) to reduce the burst loss probability. Since WCs are expensive and still immature technologically, it is desirable to reduce the number of WCs in the network and to have partial wavelength conversion capability at the core nodes. Nevertheless, a majority of algorithms in the literature are proposed under the full wavelength conversion assumption. As a result, they do not consider the burst loss caused by insufficient WCs, i.e., bursts dropped due to the unavailability of free WCs to convert them to unused wavelengths. In this paper, we demonstrate how to use burst rescheduling to decrease the burst loss due to insufficient WCs and hence cut down on the overall burst loss probability in OBS networks. Two burst rescheduling algorithms are proposed. Their effectiveness in reducing the overall burst loss probability is verified through simulation experiments.     

Dynamic queue level control of TCP/RED systems in AQM routers

One main TCP congestion control objective is, by dynamically adjusting the source window size according to the router queue level, to stabilize the buffer queue length at a given target, thereby achieving predictable queueing delay, reducing packet loss and maximizing link utilization. One difficulty therein is the TCP acknowledging actions will experience a time delay from the router to the source in a TCP system. In this paper, a time-delay control theory is applied to analyze the mechanism of packet-dropping at router and the window-updating in TCP source in TCP congestion control for a TCP/RED dynamic model. We then derive explicit conditions under which the TCP/RED system is asymptotically stable in terms of the instantaneous queue. We discuss the convergence of the buffer queue lengths in the routers. Our results suggest that, if the network parameters satisfy certain conditions, the TCP/RED system is stable and its queue length can converge to any target. We illustrate the theoretical results using ns2 simulations and demonstrate that the network can achieve good performance and converge to the arbitrary target queues.     

A network rate management protocol with TCP congestion control and fairness for all

Our study is motivated by the need to enable quality of service (QoS), congestion control and fair rate allocation for all end applications. We propose a new approach to address these needs which is different from the current practice whereby end applications pursue their own rate control using TCP. Our approach comprises a network rate management protocol (RMP) that controls the rate of all flows (at an aggregate level based on routes) subject to QoS requirements. The RMP control also facilitates a new TCP sliding-window congestion control based on the fair target rates computed by the RMP. Each non-TCP aggregate flow is policed by its respective edge router and each TCP flow adapts its window size as to achieve the RMP suggested fair target rate. The stability analysis of the new TCP congestion control is performed in a linearly scalable framework, which is less restrictive than a fluid model. We show that our proposed control is linearly scalable and establish its global asymptotic stability under arbitrary and variable information time lags, aka totally asynchronous conditions. The stability and the vitality of our control is verified by two means. One is a simulation of a network comprising 74 core links and up to 768 flows, each using its own access link. The simulation is also used to compare our control with the congestion control algorithms used in Fast, Vegas and Reno TCPs. The second verification means is an actual implementation of the control in the Linux kernel and its experimentation in a WAN testbed network comprising six routers and long haul links running UDP flows as well as CUBIC, N-RENO and C-TCP flows. Our experiments demonstrate that our approach can guarantee fair rates for all flows and QoS to premium flows.