paper digest

congestion control

[1]Mehrotra S, Li J, Sengupta S, et al. Hybrid window and rate based congestion control for delay sensitive applications[C]//Global Telecommunications Conference (GLOBECOM 2010), 2010 IEEE. IEEE, 2010: 1-5.
 window based congestion control leads to bursty network traffic which causes additional queue delay. The author proposed a method that hybrid use queue delay and packet loss to control packets rate sending to network. The goal is to develop a protocol which minimizes queuing delay and packet loss while fully utilizing link capacity and maintaining fairness across flows.The queue delay is measured as relative one way delay.Although the one way delay is sensitive to clock offset, the relative relative one way delay does not suffer this problem.
$ROWD=OWD-OWD_{min}$.
[2]Loguinov D, Radha H. End-to-end rate-based congestion control: convergence properties and scalability analysis[J]. IEEE/ACM Transactions on Networking, 2003, 11(4): 564-577.
 current real time streaming typically rely on rate based flow control. This work proposes an AIMD rate flow control derived from window based congestion control.
[3]Carofiglio G, Gallo M, Muscariello L. Optimal multipath congestion control and request forwarding in information-centric networks: Protocol design and experimentation[J]. Computer Networks, 2016, 110: 104-117.
[4]Rejaie R, Handley M, Estrin D. RAP: An end-to-end rate-based congestion control mechanism for realtime streams in the Internet[C]//INFOCOM’99. Eighteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE. IEEE, 1999, 3: 1337-1345.
 RAP uses an AIMD algorithm.The transmission rate is controlled by inter packet gap(IPG).
$S_i=\frac{packet\_size}{IPG_i},\ IPG_{i+1}=\frac{IPG_i*C}{IPG_i+C}\tag{1}$
$\alpha=S_{i+1}-S_i=\frac{packet\_size}{C}\tag{2}$
 Where $S_i$ and $\alpha$ denote rate and step height respectively. C is a constant with a dimension of time.Upon detecting congestion, the transmission rate is decreased multiplicatively, $S_{i+1}=\beta S_i,IPG_{i+1}=\frac{IPG_i}{\beta},\beta =0.5$ .
 The rate adjustment frequency depends on feedback delay and is no more than once in per rtt. Changing the rate too often leads to oscillation and infrequent changed results in unresponsive behavior.
[5]Making Available Base-RTT for Use in Congestion Control Applications
 baseRTT is an important indicator to network bottleneck queue status.But Estimate the exact baseRRT is a problem.The later coming flows may overestimate baseRTT due to queue building-up caused by earlier flows.
 The backoff factor $\beta=\delta\frac{baseRTT}{RTT_{max}}$,where $\delta$ is a design parameter.And this factor is used to probe the minimal rtt.
[6]Fair End-to-End Window-Based Congestion Control
 The work demonstrates that the existence of congestion control protocols that converge to a fair equilibrium without the help of the internal network nodes
[7]Bhandarkar S, Reddy A L, Zhang Y, et al. Emulating AQM from end hosts[C]//ACM SIGCOMM Computer Communication Review. ACM, 2007, 37(4): 349-360.
 The work proposed to emulate AQM in the end host.The main motivation behind those work are to improve the competence of dealy-base congestion control mechanism. End-host based measurements of the state of the queue essentially entails sampling the queue at certain times and has limitation of over sampling.Their observation is that (a) losses are preceded by a significant increase in RTT in very few cases, and (b) responding to a wrong prediction can result in severe degradation of performance.The high delay in network may be very short and the end host may be unable to detect it at all.
 The basic idea behind PERT is very simple and involves responding to delay in a probabilistic rather than deterministic manner.
 In total, it’s worthing reading.
[8]a transport layer approach for achieving aggregate bandwidths on multihomed mobile hosts(2002-MobiCom)
 header of line blocking, persisting timer.
[9]A compound TCP approach for high-speed and long distance networks
 Compound tcp is an congestion control mechanism implemented in Windows, which is a synergy of delay based and loss based approach.The multiplicative window reduce of loss based congestion control algorithm results low link utilization when congestion occurs.The delay based congestion control improves throughput in long delay networks.All the delay based congestion control algorithms face the potential of starvation when competing with loss based congestion control. The delay based congestion control can detect queue building-up early.The ctcp can rapidly increase sending rate when network is under utilization and maintains bandwidth fairness with tcp.
[10]Analysis of the increase and decrease algorithms for congestion avoidance in computer networks
 Chiu and Jain has shown that that the AIMD principle of TCP stack converges to a fair and equilibrium point.
 It has been observed the rate allocated to a connection is inversely proportional to round trip time. The work of Kelly has shown the rate control based on AIMD is proportional fairness instead of max-min fairness.
 The resources allocated to a particular connection will depend on to a great extent on the path through which it is routed, and the load level of the component links.
 The above sentences are cited from Computing Optimal Max-Min Fair Resource Allocation for Elastic Flows.
[11]An Analysis of AIMD Algorithm with Decreasing Increases(2004)
 The author analyses the reliable transmission protocol UDT built up on udp. The analysis theories used in this paper are base on decades of congestion control study from academia.
[12]Improved Utility-Based Congestion Control for Delay-Constrained Communication(2017-ton)
 As the buffer occupation of loss based congestion control leads to queue building-up delay.Delay based congestion control may be a viable solution to delay sensitive application.But they do not perform well when sharing bandwidth with loss base algorithms.
[13]Self-clocked Rate Adaptation for Conversational Video in LTE
 The solution conforms to the packet conservation principle and uses a hybrid loss and delay based congestion control algorithm
[14]Multimedia Congestion Control: Circuit Breakers for Unicast RTP Sessions(RFC8031)
[15]Flow and congestion control for Internet media streaming applications
 TODO
[16]End-to-End Congestion Control for the Internet:Delays and Stability
 Stability analysis of utility maximization based congestion control.
[17]Non-concave network utility maximization: A distributed optimization approach
 The network utility function is modeled as a non-concave function, since it is the best description of the quality of service perceived by users with inelastic applications, such as video and audio streaming.
[18]Congestion Control for High Bandwidth-Delay Product Networks(2002-sigcomm)(xcp)
 Congestion is not a binary variable,so congestion signal should reflect the degree of congestion.The aggressiveness of the sources should be adjusted according to the delay in the feedback-loop. The dynamics of congestion control may be abstracted as a control loop with feedback delay.
The router can get the cwnd and rtt information of a flow.
[19]High Speed Networks Need Proactive Congestion Control(2015)
 PERC computes link bandwidth share based on flows.The Tcp congestion control algorithms(reno,cubic) are reactive in essence, iteratively reacting to congestion signal and conservatively probing flow available bandwidth. Rate adaption in performed over several iteration. Reactive algorithms have slow convergence time and make flow completion time longer than necessary.(with simulation instances of omnet++ and ns2)
 Its parts on mathematical proof based on queue theory are worth for reference.
[20]Controlling Queuing Delays for Real-Time Communication: The Interplay of E2E and AQM Algorithms
 Keywords:Webrtc, CoDel,PIE,SFQ,FQ_CoDel
 The AQM(active queue management) algorithms were developed to control the queue buffer by dropping packets or marking them over the last two decades.The adoption of those algorithms have hampered by two main issues: 1)the queue length is not quite correlated with congestion,2)the complex configuration of their parameters has to be made.
[21]A delay-based approach for congestion avoidance in interconnected heterogeneous computer networks(1988)
$P=\frac{T^{\alpha}}{D}$.Throughput and Delay.
[22]A Positive Systems Model of TCP-Like Congestion Control: Asymptotic Results(2006-TON)
 From the viewpoints of tranmission protocol designer,the fllowing properties should be considered:1)network fairness;2)network convergence and responsivenss;3)network throuphput.
 Any new protocol operating the AIMD variants that satisfies $\alpha=2(1-\beta)$ will be friendly to legacy TCP.
 The total throuphput of the bottleneck when congestion occures is given by:
$R(k)^-=\frac{\sum_{i}^{n}w_i(k)}{T_d+q_{max}/B}\tag{7}$
 After window backoff,the throughput is the following under the assumption the queue buffer will be emptied:
$R(k)^+=\frac{\sum_{i}^{n}\beta_{i}w_i(k)}{T_d}\tag{8}$
 But if the sources backoff too much,data throughput will degrade and the link operates under its maximum rate. In order to gunrantee the maximum rate,the both rates should be euqated, which may be achieved by the choice of $\beta_i$.
$\beta_{i}=\frac{T_d}{T_d+q_{max}/B}=\frac{rtt_{min}}{rtt_{max}}\tag{9}$
 For unsynchronized network,when congestion event happens,(the source i congestion window is $w_i(k)$,and the window of last rtt is $w_i(k)-\alpha_{i}$):
$\sum_{i}^{n}\frac{w_i(k)-\alpha_{i}}{rtt_i,{max}}=B\tag{14}$
$W(K+1)=AW(k)\tag{5}$
The fomular (5) can be modeled by markovian chain.(please take 《Stochastic equilibria of AIMD communication networks》for more reference.)
[23] TCP拥塞控制算法纵横谈-Illinois和YeAH

Aggregate congestion control

[1]Aggregate Congestion Control for Distributed Multimedia Applications(2004)
[2]ROSIEE: Reduction of Self Inflicted Queuing Delay in WebRTC(2017)

packet schedule

[1] Load balancing over multipaths using bandwidth-aware source scheduling
 fast path first,not to much.
[2]Improved Data Distribution for Multipath TCP Communication
 This work proposes a method named Arrival-Time matching Load-Balancing(ATLB) to guarantee in order packet delivery to receiver.
$score_i=\frac{Q_i}{G_i}+\frac{srtt_i}{2}\tag{1}$
 Where $Q_i$ is the queue length at the sender’s transmitting buffer,$G_i$ is a smoothed throughput of a connection.
[3]A Unified Approach to Congestion Control and Node-Based Multipath Routing(2009-ton)
 The work propose to split traffic at immediate nodes base on routing delay price.The main idea is to increase routing traffic in links with lower than average price and decrease the rest.
[4]Delay-based Congestion Control for Multipath TCP
 At equilibrium point,all the paths used by a flow will eventually have the same price.

Utility function based multipath flow control

[1]Optimal flow control and routing in multi-path networks(2006)
 The utility function flow optimization theory constructed by Kelly and Low is quit suitable for elastic flow.It has drawback when dealing with resource allocation and flow control if real time application involved. So this work modified the utility function to make it suitable for real time application.
(2)Utility max–min fair resource allocation for communication networks with multipath routing(2009)
[2]Fairness comparison of FAST TCP and TCP Reno(2007)
[3]Non-convex optimization and rate control for multi-class services in the Internet(2005-ton)
 Concave utility functions are appreciate for modeling traditional data service,namely elastic traffic and do not capture the characteristics of inelastic service such as audio and video service.

path selection

[1]Path selection and multipath congestion control(2007-infocom)
 Analysis path selection based on stochastic theory and utility function.

machine learning and network

[1]Traffic Allocation Control using Support Vector Machine in Heterogeneous Wireless Link Aggregation(2011)