V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....links characterized by a high bit error rate. Several schemes have been proposed to decouple the congestion control of TCP from its error control. They aim to let TCP retransmit the errors without reducing considerably the throughput. TCP uses two algorithms, slow start and congestion avoidance [4], to adapt its window to the network state. We assume that W is measured in segments not in bytes. Starting at W = 1, slow start is used to increase W by one segment for every incoming ACK until we reach a threshold W th considered as an estimation of the network capacity. This exponential growth ....

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM Sigcomm'88, Stanford, CA, USA, Aug. 1988.

....sequence number (the original plus three Duplicate ACKs) indicates the loss of the packet following the last packet acknowledged. The number four is chosen to minimize the probability that a reordering of packets causes a wrong error detection. Congestion control has been added to TCP by Jacobson [26]. An additive increase multiplicative decrease strategy has been adopted to change the window size, and therefore the throughput, as a function of network conditions. The window used for AEow control is always taken as the minimum of the congestion window and the one advertised by the receiver. ....

....network for any extra bandwidth. The probe continues until a loss occurs. Here, the source supposes that the network is getting into congestion and it sets its estimation of the capacity to half the current size of the window. The rst version of TCP that implements congestion control, called Tahoe [26], sets the congestion window 2 A segment is the size of data in a TCP packet. RR n# 0123456789 6 Barakat, Altman, and Dabbous at this point to one and uses again slow start to reach the new estimation of the capacity. Slow starting after every loss deteriorates the performance given the low ....

V. Jacobson, iCongestion avoidance and controlj, in ACM SIGCOMM, Stanford, CA, USA, August 1988.

....beyond the maximum input rate, the more important congestion losses become. The evolution of the input rate is more and more controlled by the available bandwidth in the network and the buoeer sizes rather than by random losses. This is the assumption a congestion control protocol as TCP IP [9] makes about the network. The best performance is obtained when random losses are suOEciently rare. The condition for random losses to be negligible is that the average input rate during the bad state of the channel given by our model (i.e. when ignoring congestion losses) is much larger than the ....

V. Jacobson, iCongestion avoidance and controlj, in Proc. ACM Sigcomm'88, Stanford, CA, USA, August 1988.

....transmis pendant la phase slow start. Mots cl#s : TCP, Slow Start, Congestion Avoidance, Liens Satellites, Mod#lisations, Simulations, #valuation de Performance. Improving TCP over Satellite Links 3 1 Introduction TCP, the window based reliable Transmission Control Protocol of the Internet [11, 18], uses two algorithms Slow Start and Congestion Avoidance to control the AEow of packets in a network. With slow start, the window of the sender is set initially to one segment and it is increased by one for every acknowledgment (ACK) received. The result is an exponential growth of the connection ....

....this linear growth, which is slower than the exponential one during slow start, continues until we reach the maximum reachable window Wmax . This is the maximum number of TCP packets that can be t into the pipe and the buoeer. Here, a loss occurs due to buoeer overAEow. According to TCP tahoe [11] used in this paper, W th is set to half Wmax and W to 1. A new slow start and then a new TCP cycle begins. We give Wmax the value found in [7] Wmax = B Gamma ( Gamma ) 1 (3) We note here that the addition of the exogenous traOEc reduces the share of the TCP connection in the buoeer and ....

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM Sigcomm'88, Stanford, CA, USA, Aug. 1988.

....to Sr. Because all the nodes between N and D receive packets at a rate slower than their available bandwidth, their parameters don t aoeect the performance of TCP. The following analysis focuses on the impact on TCP of nodes between Sr and N . According to TCP congestion control algorithms [5, 9], if ACKs are not lost nor delayed, an ACK triggers maximum the transmission of a burst of two packets. This happens upon every ACK reception during SS and when the congestion window W increases by one segment during CA. If the nodes between Sr and N have an available bandwidth 2, the ....

....to other values of W th . A condition to not encounter the problem of losses during SS has been calculated in [1, 7] This condition accounts only for the main bottleneck. In this section, we recalculate this condition with our general model. As in these works, we consider a Tahoe version of TCP [5, 4] where SS is frequently called. We divide a SS phase into mini cycles of duration T [7] W doubles every T . During mini cycle n, Sr sends a burst of 2 n packets at rate 2. Mini cycle n 1 starts when the ACK for the rst packet of this burst reaches Sr. These bursts propagate from node to node ....

V. Jacobson, iCongestion avoidance and controlj, ACM Sigcomm, Aug. 1988.

....of our general model and then we identify their parameters from real TCP traces. 1 Introduction We study in this paper the performance of an additive increase multiplicativedecrease AEow control protocol. This is the kind of control used by TCP, the widelyused transport protocol of the Internet [9]. TCP is used as a reference through the present work, however we anticipate that our results will be also applicable for other AEow control mechanisms. A AEuid approach is used to model the controlled AEow. The transmission rate of the source is assumed to grow linearly at a rate ff. In the case ....

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

....in this paper the tradeooe between the bandwidth consumed by FEC and that gained by TCP. 1 Introduction Forward Error Correction (FEC) is widely used to improve the quality of noisy transmission media as wireless links [2, 4] This improvement is of importance for a transport protocol as TCP [8, 16] which uses the loss of packets as an indication of network congestion. A TCP packet corrupted while crossing a noisy link is discarded before reaching the receiver which results in an unnecessary window reduction at the TCP source, and hence in a deterioration of the performance of the TCP ....

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug. 1988.

....to adapt to the varying network conditions often use decentralized control protocols in order to optimize their own performances. This is, for example, the case when transmitting data over the Internet, where users themselves determine their transmission rates (typically using the TCP IP protocol [16]) The need for decentralized distributed individual controls in telecommunication networks has stimulated a substantial amount of research using game theoretic methods, in both routing and AEow control [2, 3, 6, 7, 8, 9, 11, 12, 15, 17, 18, 20, 19, 22, 23, 24, 26] Within the design of AEow ....

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM `88, Stanford, CA, 314329, August 1988.

....and characterize the parameters of a connection. Much of the work described in the literature has focused on analyzing and estimating characteristics of connections such as end to end capacity, delay, and loss. The results have been used to design of routing and AEow control algorithms (e.g. [14, 16, 5]) to dimension buoeers and link capacity (e.g. 21, 9] and to choose parameters in simulation and analytic studies (e.g. 7, 29] They have also proved very important in the design of proper mechanisms for real time applications. For example, estimating shape of the delay distribution is ....

....by including a dependence on packet length in the retransmission timeout algorithm. Several other studies have since addressed timeout adjustment in TCP, and they have proposed improvements to take into account packet losses, packet retransmissions, and the variance of packet round trip delays [14]. More recent studies have considered the delay and loss behavior in the NSFNET, and more generally in the Internet. It is convenient to classify these studies depending on the dioeerent time scales over which the above behavior has been observed. Coarse grained NSFNET statistics obtained from ....

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM Sigcomm `88, Stanford, CA, pp. 314-329, August 1988.

....is conducted over the real Internet and a comparison is provided with other models which make simple assumptions on the inter loss time process. 1. INTRODUCTION We analyze in this paper the performance of TCP (Transmission Control Protocol) the widely used transport protocol of the Internet [15, 30]. TCP is a reliable window based AEow control protocol where the window is increased until a packet loss is detected. Here, the source assumes that the network is congested and reduces its window. Once the lost packets are recovered, the source resumes its window increase. As a performance ....

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

....show that the time average of the transmission rate increases with the burstiness of losses. We nally examine the impact of burstiness of losses on the transmission rate variability. 1. INTRODUCTION Flow control mechanisms in the Internet, particularly those of the Transmission Control Protocol [12], use the loss of packets as an indication of network congestion. In general, the transmission rate of the controlled AEow is linearly increased until a loss occurs. The network is supposed here to be congested and the transmission rate is multiplicatively decreased in order to alleviate this ....

....The transmission rate is halved and the linear growth is then resumed. This model approximates the performance of several versions of TCP IP where the transmission rate at any instant is equal to the window size divided by the RTT and where the window increases linearly by 1 packet every RTT [12]. If the delay ACKs mechanism is implemented at the destination, the increase in TCP window is by one packet every two RTTs. This linear window increase corresponds to the congestion avoidance mode of TCP. The slow start mode is neglected in this paper due its fast exponential window increase. The ....

[Article contains additional citation context not shown here]

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

....in the heterogeneity of the Internet. High speed links (Fiber optic) long and variable delay paths (Satellite links) lossy links (Wireless networks) asymmetric paths (Hybrid satellite networks) and others, are becoming widely embedded in the Internet. Many works have studied by experimentations [5, 6, 10, 12], analytical modeling [3, 8, 13, 14] and simulations [7, 8, 9, 11] the performance of TCP in this new environment. Most of these works have focused on a particular environment: satellite networks, mobile networks, etc. They have shown some problems in the operation of TCP. Long propagation delay ....

....we present the proposed solutions as well as our comments in four corresponding sections. 2 Overview of TCP TCP is a reliable window based ACK clocked AEow control protocol. It uses an additive increase multiplicative decrease strategy for changing its window as a function of network conditions [1, 12]. Starting from one packet, or a larger value as we will see later, the window is increased exponentially by one packet for every non duplicate ACK until the source estimate of the network capacity is reached. By capacity of the network, sometimes called the pipe size, we mean the maximum number ....

[Article contains additional citation context not shown here]

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug. 1988.

....(i.e. the application and the network) in which this module is used. For each packet that triggers starting the retransmission timer, either an acknowledgement will be received (NA event) or the timer will go ooe (ALARM event) If the timer works correctly (i.e. does not go ooe too early [Jac88]) each time out indicates a packet loss. Thus, the probability of an ALARM event is (almost) equal to the probability of a packet loss on the network that is used. Dioeerent network types have dioeerent packet loss rates, and also loose packets for dioeerent reasons. One possible cause of packet ....

....loss on the network that is used. Dioeerent network types have dioeerent packet loss rates, and also loose packets for dioeerent reasons. One possible cause of packet loss is that the packet was damaged during network transmission. For iwiredj networks, this is usually an infrequent event ( 1 [Jac88]) For wireless networks, in contrast, packet losses occur very frequently (say 20 ) The fact whether the application (and thus the building block contained used by the application) is more often used on a wired or on a wireless network can be easily determined by the application programmer. The ....

V. Jacobson. iCongestion Avoidance and Controlj. SIGCOMM '88, 314-329.

....analysis of this rst phase as a function of the network and TCP parameters. Among our results, we show that as claimed, the previous works improve the performance on paths with large buoeers. However, on paths with small buoeers, completely dioeerent results can be obtained. 1: Introduction TCP [9, 12] is the transport protocol of the Internet. Using a congestion window (W ) it controls the AEow of application packets as a function of network congestion. W represents the maximum number of packets the source can transmit without the receipt of any acknowledgment (ACK) from the destination. At ....

....packets as a function of network congestion. W represents the maximum number of packets the source can transmit without the receipt of any acknowledgment (ACK) from the destination. At the beginning of a TCP connection, a Slow Start phase (SS) is called to increase W quickly and smoothly [9]. It is followed by a slower increase phase called Congestion Avoidance (CA) 9] The switching from SS to CA happens at a window called the SS threshold (W th ) which is the source estimate of the network capacity. By pipe size or by network capacity we mean in the sequel the maximum number of ....

[Article contains additional citation context not shown here]

V. Jacobson, iCongestion avoidance and controlj, ACM Sigcomm, Aug. 1988.

....buoeers. However, on paths with small buoeers, completely dioeerent results could be obtained. 1 Introduction TCP is the main responsible for the stability of the Internet. By varying the size of its window (W ) it controls the AEow of application packets so as to avoid network congestion [10]. The two main algorithms used by TCP for congestion control are Slow Start (SS) and Congestion Avoidance (CA) 10, 13] SS is used to increase quickly W until a certain estimate of the network capacity. By network capacity or pipe size, we mean in the sequel the maximum number of packets that can ....

....TCP is the main responsible for the stability of the Internet. By varying the size of its window (W ) it controls the AEow of application packets so as to avoid network congestion [10] The two main algorithms used by TCP for congestion control are Slow Start (SS) and Congestion Avoidance (CA) [10, 13]. SS is used to increase quickly W until a certain estimate of the network capacity. By network capacity or pipe size, we mean in the sequel the maximum number of packets that can be t on the path. The network capacity estimate is called the SS threshold (W th ) Once W th is reached, the source ....

[Article contains additional citation context not shown here]

V. Jacobson, iCongestion avoidance and controlj, ACM Sigcomm, Aug. 1988.

....of information from a source to a destination so as to avoid unacceptably long delays, high loss rates and other eoeects of congestion which are undesirable for both the individual user as well as the entire network. Examples of such controls are the TCP IP congestion control in the Internet [13], and the control of ABR (Available Bit Rate) traOEc in ATM (as was standardized by the ATM forum [1] Such control mechanisms are especially used in regulating ibest eoeortj type traOEc, such as data transfer; this type of traOEc is often Paper submitted for an invited session at the 26th ....

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM `88, Stanford, CA, 314-329, August 1988.

....of losses, and not just to the average loss rate: for a given average loss rate, we show that the time average of the transmission rate increases with the burstiness of losses. 1 Introduction Flow control mechanisms in the Internet, particularly those of the Transmission Control Protocol (TCP) [19,24], use the loss of packets as an indication of network congestion. In general, the transmission rate of the controlled AEow is linearly increased until a loss occurs. The network is supposed here to be congested and the transmission rate is multiplicatively decreased in order to alleviate this ....

....growth is then resumed. This model approximates the performance of several versions of TCP. Indeed, the transmission rate of a TCP connection at any instant is equal to the window size divided by the RTT. The window in turn increases by 1 packet for every window s worth of Acknowledgments (ACK) [19]. This results in a linear rate increase by a factor ff = 1=RTT 2 if ACKs are not delayed at the destination and ff = 1=2RTT 2 if ACKs are delayed [24] the linear growth is known to hold for TCP connections in which the bandwidth delay product is large in comparison with queueing delays, ....

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

....link at 2 Mbps) and send requests and acknowledgements (ACK) via a slow reverse channel (e.g. a dial up modem line at 64 kbps) Figure 1 shows an example of such asymmetric path. It has been shown [1, 3, 5, 8, 13] that the slowness of the reverse channel limits the throughput of TCP transfers [10, 15] running in the forward direction. A queue of ACKs builds up in the buoeer at the input of the reverse channel (we call it the reverse buoeer) causing an increase in the round trip time (RTT) and an overAEow of the buoeer (i.e. loss of ACKs) This results in a performance deterioration for many ....

....buoeer is large and that ACKs are not ltered. The window at the sender grows then exponentially with a rate function of the frequency at which the receiver acknowledges packets. Recall that we are working in the slow start mode where the window is increased by one packet upon every ACK arrival [10]. Suppose that the receiver acknowledges every d packets, thus the window increases by a factor ff = 1 1=d every RTT. Note that most of TCP implementations acknowledge every other data packet t RTT t 2RTT 3RTT W(n 1) d ACKs W(n 2) d ACKs W(n) d ACKs Time : Start of service of the W(n) d ....

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

....consisting of a single node with a number of input sources. We used Large Deviation techniques to investigate the queueing behaviour of the network under dioeerent AEow control regimes. 1 Introduction The rst documented congestion collapse of the Internet took place in October 1986 in USA [J88]. It was noticed that the data throughput at the interconnection between UC Berkeley and Lawrence Berkeley Laboratory dropped by a factor thousand. The complexity of the Internet in 1986 was far from todays and the implications of a similar collapse now would be severe and costly. The collapses ....

....for the adjustment of the service rate in the source. After extensive experimenting in search for an algorithm that was simple and reliable enough to be implemented in the TCP window control, the research community chose one of the so called additive increase multiplicative decrease type [CJ89] [J88]. In the algorithm, the AEow controlled sources adjust their traOEc rate s according to the following rule: s(t 1) ae s(t) b; no congestion ds(t) congestion where 0 d 1 and b 0. It has been shown to satisfy suOEcient conditions for convergence to an eOEcient and fair state for ....

[Article contains additional citation context not shown here]

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM Sigcomm '88, page 314-329, 1988.

....and to control the rate at which it sends data into the network accordingly. The feedback information can be implicit or explicit. In the Internet TCP, the feedback is implicit. There, the source uses timeouts and duplicate acknowledgements to detect buoeer overAEows in the intermediate nodes [14]. In the packet pair scheme, the source sends pairs of packets close together; the interarrival time between the acknowledgement packets of a pair of data packets yields a (inverse) measure of the bottleneck service rate [20] The feedback is explicit in the congestion bit scheme [31] and its many ....

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM SIGCOMM `88, Stanford, CA, pp. 314-329, August 1988.

....in order not to congestion the network and swamp all the network resources. Internet collapse could happen for example if a large number of receivers continue to subscribe to the minimal quality AEow (i.e. no more rate decrease in case of high packet loss observed) To follow the behavior of TCP [18], the application can wait for a delay T r before probing the network state by joining the rst layer. T r = 2 k oe, where k is the number of consecutive failed attempts, and oe a random value added to avoid synchronization of receivers. 4 Experimental evaluation over the MBone The layer ....

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM Sigcomm `88, Stanford, CA, pp. 314-329, Aug. 1988.

....resources are shared fairly between different users. This regulation is typically done using feedback information about the state (more or less congested) of the network. In TCP, the state of the network is characterized by packet losses and the regulation is done using a dynamic window scheme [11]. The speci cs of the control mechanism are as follows. Packets are assigned increasing sequence numbers. The source TCP sends in each packet continuous data octets accompanied by the sequence number of the rst octet. The destination TCP maintains a set of continuous sequence numbers. When it ....

.... th ) N 2 = W 2 Gamma W t 2 2 ; and R 2 = T ; T 3 = 4W 2 th Gamma tW 2 2( Gamma ) N 3 = 4W 2 th Gamma W 2 2 ; and R 3 = 1 Gamma ( 2 3 4W 2 th 2W th W W 2 2W th W Gamma ) Proof: We rst discuss the evolution of the window size over time [11]. If an acknowledgment arrives at time s, then W (s) W (s Gamma ) 1 if W (s Gamma ) W th (s Gamma ) slow start phase) W (s Gamma ) 1=bW (s Gamma )c otherwise (congestion avoidance phase) 7) where b Deltac denotes the integer part of the argument. If a loss is detected ....

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM Sigcomm '88 , Stanford, CA, USA, Aug. 1988.

....AEuid approximations, simulations. Abstract In order to control the AEow in TCP IP protocol, each TCP source uses a transmission window which indicates the number of packets that may be transmitted by the source but not yet acknowledged by the destination. We consider the Tahoe algorithm [7] developed by Van Jacobson for the adjustment of the window size and the computations of time out values. This paper analyses the synchronization and coupling that occur between several controlled sources using this algorithm. We restrict to two controlled sources for simplicity of presentation. 1 ....

....the simplifying assumptions fail to hold in the highly asymmetric case, and propose an improved, more complex model to handle that situation. More extensive analysis and other proposed models can be found in [5] Our analysis is restricted to the Tahoe version of TCP IP of Van Jacobson s algorithm [7]. However, the type of problems we encounter is typical to many other control protocols in telecommunications. In particular, similar systems of coupled dioeerential equations arrise in the analysis of the control of ABR (Available Bit Rate) in ATM [1] and in other versions of TCP IP (such as Reno ....

[Article contains additional citation context not shown here]

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM Sigcomm '88, Stanford, CA, Aug. 1988.

....to each user, and the user tries to minimize deviation from this allocated bandwidth. Flow control is typically performed dynamically: some feedback information on the congestion or on round trip delay is used to update the input rate. For example, in the Tahoe version of TCP IP congestion control [19], congestion is detected through losses or through time out mechanisms. In the Vegas version of TCP IP [11] the available bandwidth is also used as feedback information (it is obtained through estimation of round trip delays) More detailed queue length information may also be available as ....

....the rate may be conveyed from the switches to the sources through special information cells that are called RM (Resource Management) cells. Flow control is often performed in a decentralized way in telecommunication networks: each user controls its own AEow. This is the case in the Internet, see [19], or in some best eoeort type traOEc in ATM (the Unspeci ed Bit Rate transfer capacity, see [1] This gives rise to a noncooperative dynamic game: each user has its own objectives, but the actions of the dioeerent users inAEuence the quality of service of other users. Controllers that have been ....

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Stanford, CA, 314-329, Aug. 1988.

....In the besteoeort approach there is not a xed bandwidth allocated to the source, but instead, the resources allocated to the source at each moment depend on those currently available. This is the main type of service provided by the Internet (in particular, by the TCP IP congestion control [10]) and so is the Available Bit Rate (ABR) service type in ATM (see [1] A primary goal of the design of AEow control is to achieve some overall performance objectives for the dioeerent traOEc sources, such as guaranteeing some bounds on the throughput, delay, loss probabilities. In particular, ....

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM SIGCOMM `88, Stanford, CA, pp. 314-329, August 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, ACM Sigcomm, Aug 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Stanford, CA, 314-329, Aug. 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM, Aug 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM Sigcomm '88, Stanford, CA, Aug. 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM `88, Stanford, CA, 314-329, August 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, ACM SIGCOMM `88, Stanford, CA, 314-329, August 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, Proc. ACM Sigcomm'88, Stanford, CA, USA, Aug. 1988.

No context found.

V. Jacobson, iCongestion avoidance and controlj, in Proc. ACM Sigcomm'88, Stanford, CA, USA, Aug. 1988.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC