Research On TCP Congestion Control And Queue Management In DiffServ Network

With the increasing of network applications continuously, in particularly forthe development of multimedia tra?cs, the problem of congestion collapse that hadbeen resolved once appears with the new form again. It makes the problem of con-gestion control be continuously hotspot in the new network environment. The goalof congestion control is to make use of the network resources e?ectively and fairly,and to improve the network integrated performance and quality of service. Wehave developed the related researches aiming at the problems that the multimediatra?cs lack to congestion control and QoS.The innovation works consist of the following main components in the paper.1. We derived the standard model of TCP sending rate in the network trans-mission control protocol according to Probability Theory and Stochastic Process.Based on the Padhye's research, we loosed the assumptions that were used to derivethe primary model, and considered all possible working states of TCP, includingslow start that had been ignored before. So we obtained the formulation of theTCP sending rate, called Full model.2. We proposed FGAIMD algorithm that improved the core AIMD algorithmof TCP congestion control. According to the thought of Full model, we derivedthe formulation of the FGAIMD sending rate, confirmed the settings of controlparameters (α,β), and investigated the FGAIMD ?ow's index such as smoothness,response and friendliness. Comparing the FGAIMD ?ow with the AIMD ?ow, westudied the algorithms that in?uence to the network performance and match with queue management algorithm in the router.3. We proposed an improved active queue management algorithm in Di?Serv,called as WRIO, which is a modified version of the weighted queue size RIO-Calgorithm. The algorithm resolves the problem of absolute service di?erentiationthat RIO-C exists. The design thought that realize relative service di?erentiationis first put forward for queue management algorithm based on Di?Serv. We gavethe definitions of the absolute di?erentiation and the relative di?erentiation in thepaper, and considered that the relative di?erentiation should be the design indexof queue management algorithm in Di?Serv. The relative di?erentiation should beattached importance to the researcher such as fairness.We obtain the main results as follows:1. In the deriving course of the standard model of TCP sending rate, wemodify one of the Padhye's assumptions. If a packet drops in the transmissionwindow in a certain round, then Padhye supposed that all packets will be droppedafter this packet in this window. We think that the packet loss event is randomand discontinuous in the transmission window, and first consider the in?uence ofslow start which is one of the TCP working states. So we obtain the formulationof the TCP sending rate called as Full model. Comparing the Full model withthe Padhye model in the experiment, we analyze the errors of two models. Thesimulation results show that the Full model not only simplifies the expression, butalso enhances the predicted accuracy when the drop probability is not very high.Because the assumption of the Full model is close to the actual circumstance of thenetwork, the research results has the theoretic meaning and the applied value.2. We propose FGAIMD algorithm that improves the core AIMD algorithmof TCP congestion control. We introduce the control parameter (α,β) in theAIMD algorithm and get the formulation of the FGAIMD sending rate accordingto the deriving thought of Full model. Considering TCP-friendliness as the fairnesstarget, we define the error function of the FGAIMD ?ow and the AIMD ?ow. Wecalculate the optimized value that makes error minimum with the deriving model when the value ofβis given. Therefore we get a group of setting parameters thatrealize that the FGAIMD ?ow can friendly share bandwidth with the AIMD ?ow.We have validated above model and the parameter settings by the simulations.3. We investigate the FGAIMD ?ow's index such as smoothness, response andfriendliness. The results show as follows. (1) With the increasing ofβ, the sendingrate of the FGAIMD ?ow ?uctuates small down, so the algorithm can improve thesmoothness. (2) The response speed that responds to congestion relates with theβvalue, the response speed of the ?ow with the biggerβis slower; it spends moretime to release the same amount of bandwidth. On the contrary, the aggressivespeed relates with theαvalue, the aggressive speed of the ?ow with the biggerαincreases quickly; it spends shorter time to obtain the same amount of bandwidth.4. We study FGAIMD algorithm to know how it a?ects the network perfor-mance parameters through the experiment. The results show as follows. (1) InRED, regardless of light or heavy network loads, as long as the FGAIMD ?owsadopt TCP-friendly setting parameters, di?erent setting parameters or di?erentsource algorithms almost have no the in?uence on drop probability and through-put; but in Droptail, the drop probability of the AIMD ?ows are higher, and theoccupied bandwidth are lower. (2) In RED, regardless of light or heavy networkloads, the average queue sizes are shorter, and are e?ectively controlled betweenthe maximum and the minimum thresholds. The queue sizes ?uctuate less, andthey become shorter with the increasing of parameters ofβ. But in Droptail, theaverage queue sizes increase very much with the augment ofβ, and the queuesizes ?uctuate too big. In a word, the network performance is the best when theFGAIMD source algorithm cooperates with RED router algorithm.5. We investigate the performance of RIO-C algorithm by the simulations,and find that the algorithm di?erentiates the priority of packets absolutely. It willreduce the link utilization. The absolute di?erentiation is that the packets withthe higher priory are to be dropped when the packets with the lower priory almostall are dropped. It should result in continuously dropping packets of the ?ow withthe lower priory and decreasing the network utilization. Its in?uence is obvious to the TCP ?ows. The paper proposes a new active queue management algorithmcalled WRIO. We give the design of WRIO, and analyze the priory di?erentiationin the category. The simulations validate that the algorithm realizes the relateddi?erentiation and improves the link utilization. We analyze the rationality of thethreshold settings in WRIO compared with RIO-C. By the records of the randomdrop and the forced drop to red packets, we verified the accuracy of the thresholdanalysis.Our works need further studying and perfecting, such as followings. (1) De-sign adaptive parameter settings to (α,β) of the FGAIMD congestion control.(2) Study the scientific measure index of the related di?erentiation for the activequeue management algorithm in Di?Serv. (3) Resolve the problems of fairness andparameter setting for WRIO algorithm. (4) Verify and realize the algorithms inpractical Internet network.