Chiu, D., and Jain, R. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks 17, 1 (June 1989), 1-14.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... is based upon the output rate ( i ) and not the input rate ( i ) We therefore refer to our policy as AIMD ER (Additive Increase and Multiplicative Decrease using Egress Rate) We retain the multiplicative aspect in the decrease policy because it promotes fairness as shown by Chiu and Jain [4]. Furthermore, a schematic of our scheme s steady state input rate dynamics (Figure 2) looks di erent from an equivalent one for TCP or other AIMD schemes. Speci cally, the schematic for AIMD schemes resembles a sawtooth whereas our scheme holds a constant rate while the queue drains. We ....

Chiu, D., and Jain, R. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks 17, 1 (June 1989), 1-14.

....Internet, it is unclear how well each scheme performs when congestion is persistent. In addition, by modifying the linear increase multiplicative decrease algorithm of TCP, these modifications cannot ensure that max min fair sharing occurs between connections which are multiplexed across the link [8, 38]. With the exception of Tri S and Vegas, one of the problems with the TCP congestion control algorithm over current networks is that the sending sources reduce their transmission rates only 8 min th max th Dropping Marking Probability 0 Average Queue Length 1 max p Figure 2.3: The ....

....of TCP s current linear increase and a calculated bandwidth based increase. Note that one of the disadvantages of the bandwidth based increase algorithm is that it falls in a class of algorithms which have been theoretically shown to be unable to provide max min fair sharing in a responsive manner [8, 38]. While fairness is a concern, even with TCP s current congestion control algorithm, fairness between connections has already been shown to be poor in times of congestion and among connections with varying round trip times [25, 48] In addition, the idealized model used in [8] assumes that ....

[Article contains additional citation context not shown here]

D. Chiu and R. Jain. Analysis of the Increase/Decrease Algorithms for Congestion Avoidance in Computer Networks. Journal of Computer Networks and ISDN, 17(1), June 1989.

.... implementation requirements of our scheme in between purely implicit congestion detection schemes (eg: TCP [11] and explicit feedback schemes ( 25] 6] 15] 19] Moreover, our rate increase decrease policy differs from the well known additive increase multiplicative decrease (AIMD) policy [2]. We call our increase decrease policy AIMD ER. The steady state dynamics are demonstrated in Figure 2. In particular, the multiplicative decrease factor is applied to the explicitly fed back output rate ( ER ) and not to the input rate limit as in AIMD. This allows us to choose a backoff ....

CHIU, D., AND JAIN, R. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks 17, 1 (June 1989), 1--14.

.... is based upon the output rate ( i ) and not the input rate ( i ) We therefore refer to our policy as AIMD ER (Additive Increase and Multiplicative Decrease using Egress Rate) We retain the multiplicative aspect in the decrease policy because it promotes fairness as shown by Chiu and Jain [4]. Furthermore, a schematic of our scheme s steady state input rate dynamics (Figure 2) looks di erent from an equivalent one for TCP or other AIMD schemes. Speci cally, the schematic for AIMD schemes resembles a saw tooth whereas our scheme holds a constant rate while the queue drains. l maxmax ....

Chiu, D., and Jain, R. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks 17, 1 (June 1989), 1-14. 19

....when congestion is no longer present can collectively be called the congestion control aspect of the protocol. The majority of the trac on the Internet is controlled by the TCP protocol. TCP implements several congestion control mechanisms that are based on the guidelines rst proposed in [6]. TCP has been instrumental in maintaining the stability of the Internet, leading to its explosive growth in the last few years. The principle driving force behind this growth has been the development of the Web, with its inherently multi media nature. With increasing availability of high ....

.... slowly (i.e. by one packet per round trip time, which is an additive increase) When an incoming ACK indicates a lost packet, the sender assumes that the packet loss was caused by congestion in the network, and reduces its sending rate by 50 , i.e. multiplicative decrease) It has been shown in [6] that this additiveincrease, multiplicative decrease mechanism results in fair sharing of bandwidth among multiple ows. If a large number of packets are dropped by the network at once, the TCP sender may not receive any acknowledgments at all, and is forced to time out. In this mode a TCP ....

Chiu, D. M., and Jain, R. Analysis of the Increase/Decrease Algorithms for Congestion Avoidance in Computer Networks. Computer Networks and ISDN Systems 17 (June 1989).

....designed to nd the maximal rate at which TCP can send packets under current conditions without incurring packet drops. While the AIMD approach is widely considered the most appropriate one largely based on its empirical success and on certain control theoretic arguments of Chiu and Jain [1] here we seek to broaden our understanding of congestion probing algorithms in a more algorithmic direction. We formulate congestion probing as an optimization problem described in Section 2. We rst introduce, in Section 2.1, the case when the other trac remains constant, which turns out to be a ....

D. Chiu and R. Jain. Analysis of the Increase/Decrease Algorithms for Congestion Avoidance in Computer Networks. In Journal of Computer Networks and ISDN, 17(1):1-14(1989).

.... is based upon the output rate ( i ) and not the input rate ( i ) We therefore refer to our policy as AIMD ER (Additive Increase and Multiplicative Decrease using Egress Rate) We retain the multiplicative aspect in the decrease 2 policy because it promotes fairness as shown by Chiu and Jain [3]. Furthermore, a schematic of our scheme s steady state input rate dynamics (Figure 2) looks di erent from an equivalent one for TCP or other AIMD schemes. Speci cally, the schematic for AIMD schemes resembles a sawtooth whereas our scheme holds a constant rate while the queue drains. We ....

Chiu, D., and Jain, R. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks 17, 1 (June 1989), 1-14.

....service class, and each service class has an associated rate weight. A flow is assigned the rate weight associated with its class. B. Linear Increase Multiplicative Decrease In wireline networks, the linear increase multiplicative decrease (LIMD) algorithm is predominantly used for rate control [6]. LIMD is usually implemented as a sender based rate control algorithm that is based on end to end loss feedback. Let us consider a framework in which periodic feedback is used, and let r i denote the sending rate maintained by the sender during epoch i, f i denote the loss feedback from the ....

....intuition than a rigorous treatment of the acceptable operation range of the algorithm. VI. RELATED WORK AND SUMMARY We place our work in the context of related work in two areas: rate control for wireline and wireless networks, and service differentiation. The seminal work of Chiu and Jain in [6] showed that the LIMD rate control paradigm is the only robust approach for achieving convergence to fairness from any arbitrary starting configuration. It has thus formed the basis of most currently deployed rate control algorithms that are deployed today, including the different flavors of TCP. ....

D. Chiu and R. Jain. Analysis of the Increase/Decrease Algorithms for Congestion Avoidance in Computer Networks. Journal of Computer Networks and ISDN, vol. 17, no. 1, June 1989.

....and distributing detailed information about the bandwidth share a sender should use in order to avoid congesting the network. Therefore, a popular approach for heuristically determining the appropriate transmission rate is the additive increase and multiplicative decrease (AIMD) approach [4]. With this approach the sender increases its transmission rate by an additive value after receiving feedback information from the receiver indicating an underload situation in the network and decreases it multiplicatively otherwise. Instead of introducing a new communication protocol for ....

....in which it oscillates around the optimal value. The time taken to reach this equilibrium (responsiveness or transient state) and the size of the oscillations (smoothness) jointly determine the convergence. Ideally, the response time as well as the oscillation should be small, see Fig. 1 and [4]. Goal Resource Allocation Time Smoothness Responsivness Figure 1. Responsiveness and smoothness of AIMD schemes 4.1. Options for Designing the Increase Phase 4.1.1 Constant Additive Increase Rate With this approach the sender increases its transmission rate during underload periods by ....

[Article contains additional citation context not shown here]

D. Chiu and R. Jain. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks and ISDN, 17(1):1--14, June 1989.

.... higher than the sum of the initial rates of the first and second layer (564 kbit s) The oscillations are a result of the adaptation to the bursty background traffic and are typical to the additive increase multiplicative decrease kind of adaptation scheme similar to the one we are using here [6]. Figures 12 through 14 show the data losses occurring on link 3, link 2 and link 1, respectively. The results show that with the ALT scheme, the sender adapts the transmission rate for each layer in accordance with the available network resources. Thereby, the QoS of the different receivers ....

D. Chiu and R. Jain. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks and ISDN, 17(1):1--14, June 1989.

....go down more than those with smaller shares making the allocation more fair. However, with no information about the explicit share to use, ID schemes do not converge to a single steady state. Instead, the system reaches an equilibrium (steady state) in which it oscillates around the optimal state [5]. The time taken to reach this equilibrium (responsiveness or transient state) and the size of the oscillations (smoothness) jointly determine the convergence. Ideally, the time as well as the oscillation should be small, see Fig. 1 [5] An an examGoal Resource Allocation Time Smoothness ....

....(steady state) in which it oscillates around the optimal state [5] The time taken to reach this equilibrium (responsiveness or transient state) and the size of the oscillations (smoothness) jointly determine the convergence. Ideally, the time as well as the oscillation should be small, see Fig. 1 [5]. An an examGoal Resource Allocation Time Smoothness Responsivness Figure 1: Responsiveness and smoothness of increase decrease schemes ple for a TCP friendly adaptation scheme based on the increase decrease concept, we describe here an adaptation algorithm that is based on previous work in ....

D. Chiu and R. Jain. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks and ISDN, 17(1):1--14, June 1989. 19

....as the minimum of tcp s current linear increase and a calculated bandwidth based increase. We note that one of the key disadvantage of the bandwidth based increase algorithm is that it falls in a class of algorithms which has been theoretically shown to not provide responsive max min fair sharing [3, 14]. While fairness is a concern, even with tcp s current congestion control algorithm, fairness between connections has already been shown to be poor in times of congestion and among connections with varying roundtrip times [10, 19] In addition, the idealized model used in [3] assumes that ....

.... fair sharing [3, 14] While fairness is a concern, even with tcp s current congestion control algorithm, fairness between connections has already been shown to be poor in times of congestion and among connections with varying roundtrip times [10, 19] In addition, the idealized model used in [3] assumes that congestion notification is given to all sources when the bottleneck resource becomes saturated. With more intelligent queueing algorithms such as red and fred [9, 16] that deliver congestion notification preferentially to higher bandwidth flows, it may be possible for a larger class ....

D. Chiu and R. Jain. Analysis of the Increase/Decrease Algorithms for Congestion Avoidance in Computer Networks. Journal of Computer Networks and ISDN, 17(1), June 1989.

....to a given queue or bottleneck. Also note that if the losses are randomized, C= p p is a bound on the window size for all connections through any bottleneck or sequence of bottlenecks. Furthermore, connections which share the same (randomized loss) bottleneck tend to equalize their windows [CJ89] We suspect that this is the implicit resource allocation principle already in effect in the Internet today. 13 3.4 Effect of TCP Implementation FACK RH FACK RH DA Sack1 Sack1 DA Reno Reno DA New Reno New Reno DA Tahoe Tahoe DA Proportion loss (p) 0.00 0.10 0.20 0.30 0.40 ....

D. Chiu and R. Jain. Analysis of the Increase /Decrease Algorithms for Congestion Avoidance in Computer Networks. Journal of Computer Networks and ISDN, 17(1):1-- 14, June 1989.

....is equally divided between the new and the old ows. Figure 14 gives the aggregate throughput obtained by three ows started at the same time. The aggregate throughput obtained by all the three ows with WTCP is higher than that obtained with the other mechanisms. We also used fairness index [16] given by F (x) P x i ) 2 n( P x i 2 ) where x i is the share of bandwidth obtained by a ow i, to measure the fairness. The fairness index is a continuous function in [0; 1] and larger the index value, better the fairness. All the 3 congestion control mechanisms considered were ....

D. Chiu and R. Jain, \Analysis of the increase/decrease algorithms for congestion avoidance in computer networks," Journal of Computer Networks and ISDN, Vol. 17, pp. 1{ 14, June 1989.

....algorithm used in our experiments. We note that one of the key disadvantage of bandwidth based increase algorithm is that it falls under a class of algorithms which have theoretically been shown to not provide responsive max min fair sharing independent of the starting state of the network [3, 14]. While fairness is a concern, even with tcp s current congestion control algorithm, fairness between connections has already been shown to be poor in times of congestion [19] and across multiple roundtrip times [10] In addition, the idealized model used in [3] assumes that congestion ....

.... starting state of the network [3, 14] While fairness is a concern, even with tcp s current congestion control algorithm, fairness between connections has already been shown to be poor in times of congestion [19] and across multiple roundtrip times [10] In addition, the idealized model used in [3] assumes that congestion notification is given to all sources when the bottleneck resource becomes saturated. With queueing algorithms such as red and fred [9, 16] that deliver congestion notification preferentially to higher bandwidth flows, it may be possible for a larger class of ....

D. Chiu and R. Jain. Analysis of the Increase/Decrease Algorithms for Congestion Avoidance in Computer Networks. Journal of Computer Networks and ISDN, 17(1), June 1989.

....rate, i.e. exponential (slow start) if cwnd ssthresh, linear (congestion avoidance) otherwise. If cwnd ssthresh, r fir; otherwise, r r OE. fi and OE are network parameters. Currently, the function f D is 1 T (1 ack.fraction) This results in a multiplicative backoff (as recommended in [2]) which is less reactive than TCP because the rate (and hence congestion window) is reduced by a multiplicative factor of 1=n rather than 1=2 n for n packet losses. There are four key points to note in our congestion control algorithm: a) since congestion control is based on the fraction of ....

D. Chiu and R. Jain. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks and ISDN, vol 17, no 1, June 1989.

....part from the additive increase multiplicative decrease approach and in part from the partial incorrectness of the TCP model. Additive increase multiplicative decrease schemes are inherently oscillatory. That is, such systems do not converge to an equilibrium but oscillate around the optimal state [7]. Also, the TCP model we are using for determining the maximum allowed transmission rate does not consider the bandwidth available on the network or the number of connections sharing a link. Thereby, it might result in throughput values much higher 10 100 1000 10000 0 2 4 6 8 10 12 14 16 Rate ....

D. Chiu and R. Jain. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks and ISDN, 17(1):1--14, June 1989.

....this scenario, Reno TCP performs Fast Retransmit (linear increase starting from half the current congestion window) and Tahoe TCP performs slow start. For HPF, the congestion window is adjusted based on the fraction of packets received. This results in a multiplicative backoff (as recommended in [2]) which is less reactive than TCP for large windows and small loss. Note that as a last step in determining the congestion window size, cwnd is always bounded by the minimum of the congestion window and the advertised receiver window (obtained from flow control) 3.1.3. Impact of ACKs and ....

D. Chiu and R. Jain. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. Journal of Computer Networks and ISDN, 17(1):1--14, June 1989.

No context found.

D. Chiu and R. Jain. Analysis of the Increase/Decrease Algorithms for Congestion Avoidance in Computer Networks. Journal of Computer Networks and ISDN Systems, 1989.

No context found.

D. Chiu and R. Jain. Analysis of the Increase/Decrease Algorithms for Congestion Avoidance in Computer Networks. Journal of Computer Networks and ISDN Systems, 1989.

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