Congestion Control and Traffic Management in ATM Networks:
Recent Advances and A Survey

Congestion control mechanisms for ATM networks as selected by the ATM Forum traffic management group are described. Reasons behind these selections are explained. In particular, selection criterion for selection between rate-based and credit-based approach and the key points of the debate between these two approaches are presented. The approach that was finally selected and several other schemes that were considered are described.

Table of the Contents

1. Introduction
2. Quality of Service (QoS) and Traffic Attributes
3. Service Categories
4. Congestion Control Methods
5. Selection Criteria
6. Congestion Schemes
7. Credit-Based Approach
8. Rate-Based Approach
9. Credit vs Rate Debate
10. Summary
11. Acknowledgement
12. References

1 Introduction

Traffic management is concerned with ensuring that users get their desired quality of service. The problem is especially difficult during periods of heavy load particularly if the traffic demands cannot be predicted in advance. This is why congestion control, although only a part of the traffic management issues, is the most essential aspect of traffic management. Congestion control is critical in both ATM and non-ATM networks. When two bursts arrive simultaneously at a node, the queue lengths may become large very fast resulting in buffer overflow. Also, the range of link speeds is growing fast as higher speed links are being introduced in slower networks of the past. At the points, where the total input rate is larger than the output link capacity, congestion becomes a problem. The protocols for ATM networks began being designed in 1984 when the Consultative Committee on International Telecommunication and Telegraph (CCITT) - a United Nations Organization responsible for telecommunications standards - selected ATM as the paradigm for its broadband integrated service digital networks (B-ISDN). Like most other telecommunications standards, the ATM standards specify the interface between various networking components. A principal focus of these standards is the user-network interface (UNI), which specifies how a computer system (which is owned by a user) should communicate with a switch (which is owned by the network service provider).

ATM networks are connection-oriented in the sense that before two systems on the network can communicate, they should inform all intermediate switches about their service requirements and traffic parameters. This is similar to the telephone networks where a circuit is setup from the calling party to the called party. In ATM networks, such circuits are called virtual circuits (VCs). The connections allow the network to guarantee the quality of service by limiting the number of VCs. Typically, a user declares key service requirements at the time of connection set up, declares the traffic parameters and may agree to control these parameters dynamically as demanded by the network.

This paper is organized as follows. Section 2defines various quality of service attributes. These attributes help define various classes of service in Section 3. We then provide a general overview of congestion control mechanisms in Section 4and describe the generalized cell algorithm in Section 4.1. Section 5describes the criteria that were set up for selecting the final approach. A number of congestion schemes are described briefly in Section 6with the credit-based and rate-based approaches described in more details in Sections 7and 8. The debate that lead to the eventual selection of the rate-based approach is presented in Section 9. The description presented here is not intended to be a precise description of the standard. In order to make the concept easy to understand, we have at times simplified the description. Those interested in precise details should consult the standards documents, which are still being developed as of this writing in January 1995.

2 Quality of Service (QoS) and Traffic Attributes

1. Peak Cell Rate (PCR): The maximum instantaneous rate at which the user will transmit.
2. Sustained Cell Rate (SCR): This is the average rate as measured over a long interval.
3. Cell Loss Ratio (CLR): The percentage of cells that are lost in the network due to error and congestion and are not delivered to the destination. Cell Loss Ratio = (Lost Cells) / (Transmitted Cells) Each ATM cell has a ``Cell Loss Priority (CLP)" bit in the header. During congestion, the network first drops cells that have CLP bit set. Since the loss of CLP=0 cell is more harmful to the operation of the application, CLR can be specified separately for cells with CLP=1 and for those with CLP=0.
4. Cell Transfer Delay (CTD): The delay experienced by a cell between network entry and exit points is called the cell transfer delay. It includes propagation delays, queueing delays at various intermediate switches, and service times at queueing points.
5. Cell Delay Variation (CDV): This is a measure of variance of CTD. High variation implies larger buffering for delay sensitive traffic such as voice and video. There are multiple ways to measure CDV. One measure called ``peak-to-peak'' CDV consists of computing the difference between the $(1-\alpha)$-percentile and the minimum of the cell transfer delay for some small value of $\alpha$.
6. Cell Delay Variation Tolerance (CDVT)and Burst Tolerance (BT): For sources transmitting at any given rate, a slight variation in the inter-cell time is allowed. For example, a source with a PCR of 10,000 cells per second should nominally transmits cells every 100~ $\mu$s. A leaky bucket type algorithm called ``Generalized Cell Rate Algorithm (GCRA)'' is used to determine if the variation in the inter-cell times is acceptable. This algorithm has two parameters. The first parameter is the nominal inter-cell time (inverse of the rate) and the second parameter is the allowed variation in the inter-cell time. Thus, a GCRA(100$\mu$s, 10$\mu$s), will allow cells to arrive no more than 10~$\mu$s earlier than their nominal scheduled time. The second parameter of the GCRA used to enforce PCR is called Cell Delay Variation Tolerance (CDVT) and of that used to enforce SCR is called Burst Tolerance (BT).

7. Maximum Burst Size (MBS): The maximum number of back-to-back cells that can be sent at the peak cell rate but without violating the sustained cell rate is called maximum burst size (MBS). It is related to the PCR, SCR, and BT as follows:
   
   

   Since MBS is more intuitive than BT, signalling messages use MBS. This means that during connection setup, a source is required to specify MBS. BT can be easily calculated from MBR, SCR, and PCR. Note that PCR, SCR, CDVT, BT, and MBS are input traffic characteristics and are enforced by the network at the network entry. CLR, CTD, and CDV are qualities of service provided by the network and are measured at the network exit point.

8. Minimum Cell Rate (MCR): The is the minimum rate desired by a user.

3 Service Categories

1. For all service categories, the cell loss ratio may be unspecified for CLP=1.
2. Minimized for sources that adjust cell flow in response to control information.
3. May not be subject to connection admission control and user parameter control procedures.
4. Represents the maximum rate at which the source can send as controlled by the control information.
5. These parameters are either explicitly or implicitly specified.
6. Different values of CDVT may specified for SCR and PCR.

1. Constant Bit Rate (CBR): This category is used for emulating circuit switching. The cell rate is constant. Cell loss ratio is specified for CLP=0 cells and may or may not be specified for CLP=1 cells. Examples of applications that can use CBR are telephone, video conferencing, and television (entertainment video).
2. Variable Bit Rate (VBR): This category allows users to send at a variable rate. Statistical multiplexing is used and so there may be a small nonzero random loss. Depending upon whether or not the application is sensitive to cell delay variation, this category is subdivided into two categories: Real time VBR and Nonreal time VBR. For nonreal time VBR, only mean delay is specified, while for realtime VBR, maximum delay and peak-to-peak CDV are specified. An example of realtime VBR is interactive compressed video while that of nonreal time VBR is multimedia email.
3. Available Bit Rate (ABR): This category is designed for normal data traffic such as file transfer and email. Although, the standard does not require the cell transfer delay and cell loss ratio to be guaranteed or minimized, it is desirable for switches to minimize the delay and loss as much as possible. Depending upon the congestion state of the network, the source is required to control its rate. The users are allowed to declare a minimum cell rate, which is guaranteed to the VC by the network. Most VCs will ask for an MCR of zero. Those with higher MCR may be denied connection if sufficient bandwidth is not available.
4. Unspecified Bit Rate (UBR): This category is designed for those data applications that want to use any left-over capacity and are not sensitive to cell loss or delay. Such connections are not rejected on the basis of bandwidth shortage (no connection admission control) and not policed for their usage behavior. During congestion, the cells are lost but the sources are not expected to reduce their cell rate. In stead, these applications may have their own higher-level cell loss recovery and retransmission mechanisms. Examples of applications that can use this service are email, file transfer, news feed, etc. Of course, these same applications can use the ABR service, if desired.

4 Congestion Control Methods

Figure 1shows how the duration of congestion affects the choice of the method. The best method for networks that are almost always congested is to install higher speed links and redesign the topology to match the demand pattern.

Figure 1: Congestion techniques for various congestion durations

For sporadic congestion, one method is to route according to load level of links and to reject new connections if all paths are highly loaded. This is called ``connection admission control (CAC)." The ``busy" tone on telephone networks is an example of CAC. CAC is effective only for medium duration congestion since once the connection is admitted the congestion may persist for the duration of the connection. For congestions lasting less than the duration of connection, an end-to-end control scheme can be used. For example, during connection setup, the sustained and peak rate may be negotiated. Later a leaky bucket algorithm may be used by the source or the network to ensure that the input meets the negotiated parameters. Such ``traffic shaping algorithms" are open loop in the sense that the parameters cannot be changed dynamically if congestion is detected after negotiation. In a closed loop scheme, on the other hand, sources are informed dynamically about the congestion state of the network and are asked to increase or decrease their input rate. The feedback may be used hop-by-hop (at datalink layer) or end-to-end (transport layer). Hop-by-hop feedback is more effective for shorter term overloads than the end-to-end feedback. For very short spikes in traffic load, providing sufficient buffers in the switches is the best solution.

Notice that solutions that are good for short term congestion are not good for long-term overload and vice-versa. A combination of various techniques (rather than just one technique) is used since overloads of various durations are experienced on all networks. UNI 3.0 allows CAC, traffic shaping, and binary feedback (EFCI). However, the algorithms for CAC are not specified. The traffic shaping and feedback mechanisms are described next.

4.1 Generalized Cell Rate Algorithm (GCRA)

4.2 Feedback Facilities

1. Generalized Flow Control (GFC)
2. Explicit Forward Congestion Indication (EFCI)

Figure 2: ATM Cell header format

As shown in Figure 2, the first four bits of the cell header at the user-network interface (UNI) were reserved for GFC. At network-network interface, GFC is not used and the four bits are part of an extended virtual path (VP) field. It was expected that the GFC bits will be used by the network to flow control the source. The GFC algorithm was to be designed later. This approach has been abandoned.

Switches can use the payload type field in the cell header to convey congestion information in a binary (congestion or no congestion) manner. When the first bit of this field is 0, the second bit is treated as ``explicit forward congestion indication (EFCI) bit." For all cells leaving their sources, the EFCI bit is clear. Congested switches on the path, if any, can set the EFCI bit. The destinations can monitor EFCI bits and ask sources to increase or decrease their rate. The exact algorithm for use of EFCI was left for future definition. The use of EFCI is discussed later in Section 8.

5 Selection Criteria

5.1 Scalability

5.2 Optimality

Figure 3: Sample configuration for max-min fairness

The following example illustrates the above concept of max-min fairness. Figure~ 3shows a network with four switches connected via three 150 Mbps links. Four VCs are setup such that the first link L1 is shared by sources S1, S2, and S3. The second link is shared by S3 and S4. The third link is used only by S4. Let us divide the link bandwidths fairly among contending sources. On link L1, we can give 50 Mbps to each of the three contending sources S1, S2, and S3. On link L2, we would give 75 Mbps to each of the sources S3 and S4. On link L3, we would give all 155 Mbps to source S4. However, source S3 cannot use its 75 Mbps share at link L2 since it is allowed to use only 50 Mbps at link L1. Therefore, we give 50 Mbps to source S3 and construct a new configuration shown in Figure~ 4, where Source S3 has been removed and the link capacities have been reduced accordingly. Now we give 1/2 of the link L1's remaining capacity to each of the two contending sources: S1 and S2; each gets 50 Mbps. Source S4 gets the entire remaining bandwidth (100 Mbps) of link L2. Thus, the fair allocation vector for this configuration is (50, 50, 50, 100). This is the max-min allocation.

Figure 4: Configuration after removing VC 3.

Notice that max-min allocation is both fair and efficient. It is fair in the sense that all sources get an equal share on every link provided that they can use it. It is efficient in the sense that each link is utilized to the maximum load possible.

It must be pointed out that the max-min fairness is just one of several possible optimality criteria. It does not account for the guaranteed minimum (MCR). Other criterion such as weighted fairness have been proposed to determine optimal allocation of resources over and above MCR.

5.3 Fairness Index

5.4 Robustness

5.5 Implementability

5.6 Simulation Configurations

Figure 5: Theatre parking lot

For computer networks, an $n$-stage parking lot configuration consists of $n$ switches connected in a series. There are $n$ VCs. The first VC starts from the first switch and goes to the end. For the remaining $i$th VC starts at the $i-1$th switch. A 3-switch parking lot configuration is shown in Figure~ 6.

Figure 6: Parking lot configuration

5.7 Traffic Patterns

1. Persistent Sources: These sources, also known as ``greedy" or ``infinite" sources always have cells to send. Thus, the network is always congested.
2. Staggered Source: The sources start at different times. This allows us to study the ramp-up (or ramp-down) time of the schemes.
3. Bursty Sources: These sources oscillate between active state and idle state. During active state, they generate a burst of cells [ 10]. This is a more realistic source model than a persistent source. With bursty sources, if the total load on the link is less than 100\%, then throughput and fairness are not at issue, what is more important is the ``burst response time" -- the time from ``first cell in" to ``the last cell out." If the bursts arrive at a fixed rate, it is called ``open loop.'' A more realistic scenario is when the next burst arrives some time after the response to the previous burst has been received. In this later case, the burst arrival is affected by network congestion and so the traffic model is called ``closed loop.''

6 Congestion Schemes

6.1 Fast Resource Management

The fast resource management proposal was not accepted at the ATM Forum primarily because it would either cause excessive delay during normal operation or excessive loss during congestion.

6.2 Delay-Based Rate Control

Although the proposal was presented at the ATM Forum, it was not followed up and the precise details of how the delay will be used were not presented. Also, this method does not really require any standardization, since any source-destination pair can do this without involving the network.

6.3 Backward Explicit Congestion Notification (BECN)

6.4 Early Packet Discard

Note that this method does not require any inter-switch or source-switch communication and, therefore, it can be used without any standardization. Many switch vendors are implementing it.

6.5 Link Window with End-to-End Binary Rate

6.6 Fair queueing with Rate and Buffer feedback

7 Credit-Based Approach

8 Rate-Based Approach

This particular algorithm uses a ``negative polarity of feedback" in the sense that RM cells are sent only to decrease the rate but no RM cells are required to increase the rate. A positive polarity, on the other hand, would require sending RM cells for increase but not on decrease. If RM cells are sent for both increase and decrease, the algorithm would be called bipolar. The problem with negative polarity is that if the RM cells are lost due to heavy congestion in the reverse path, the sources will keep increasing their load on the forward path and eventually overload it. This problem was fixed in the next version by using positive polarity. The sources set EFCI on every cell except the nth cell. The destination will send an ``increase" RM cell to source if they receive any cells with the EFCI off. The sources keep decreasing their rate until they receive a positive feedback. Since the sources decrease their rate proportional to the current rate, this scheme was called ``proportional rate control algorithm (PRCA)." PRCA was found to have a fairness problem. Given the same level of congestion at all switches, the VCs travelling more hops have a higher probability of having EFCI set than those travelling smaller number of hops. If p is the probability of EFCI being set on one hop, then the probability of it being set for an n-hop VC is $1-(1-p)^n$ or $np$. Thus, long path VCs have fewer opportunities to increase and are beaten down more often than short path VCs. This was called the ``beat-down problem [ 3]." One solution to the beat down problem is the selective feedback [ 31] or intelligent marking [ 1] in which a congested switch takes into account the current rate of the VC in addition to its congestion level in deciding whether to set the EFCI in the cell. The switch computes a ``fair share" and if congested it sets EFCI bits in cells belonging to only those VCs whose rates are above this fair share. The VCs with rates below fair share are not affected.

8.1 The MIT Scheme

Secondly and more importantly, the binary feedback schemes were designed for window-based controls and are too slow for rate-based contols. With window-based control a slight difference between the current window and the optimal window will show up as a slight increase in queue length. With rate-based control, on the other hand, a slight difference in current rate and the optimal rate will show up as continuously increasing queue length [ 15, 16]. The reaction times have to be fast. We can no longer afford to take several round trips that the binary feedback requires to settle to the optimal operation. The explicit rate feedback can get the source to the optimal operating point within a few round trips. The explicit rate schemes have several additional advantages. First, policing is straight forward. The entry switches can monitor the returning RM cells and use the rate directly in their policing algorithm. Second with fast convergence time, the system come to the optimal operating point quickly. Initial rate has less impact. Third, the schemes are robust against errors in or loss of RM cells. The next correct RM cell will bring the system to the correct operating point. We substantiated our arguments with simulation results for an explicit rate scheme designed by Anna Charny during her master thesis work at the Massachusetts Institute of Technology (MIT) [ 5]. The MIT scheme consists of each source sending an RM cell every nth data cell. The RM cell contains the VC's current cell rate (CCR) and a ``desired rate." The switches monitor all VC's rates and compute a ``fair share." Any VC's whose desired rate is less than the fair share is granted the desired rate. If a VC's desired rate is more than the fair share, the desired rate field is reduced to the fair share and a ``reduced bit" is set in the RM cell. The destination returns the RM cell back to the source, which then adjusts its rate to that indicated in the RM cell. If the reduced bit is clear, the source could demand a higher desired rate in the next RM cell. If the bit is set, the source use the current rate as the desired rate in the next RM cell. The fair share is computed using a iterative procedure as follows. Initially, the fair share is set at the link bandwidth divided by the number of active VC's. All VCs, whose rates are less than the fair share are called ``underloading VCs". If the number of underloading VCs increases at any iteration, the fair share is recomputed as follows:

This proposal was well received except that the computation of fair share requires order $n$ operations, where $n$ is the number of VCs. Search for an O(1) scheme led to the EPRCA algorithm discussed next.

8.2 Enhanced PRCA (EPRCA)

The RM cells contain desired explicit rate (ER), current cell rate (CCR), and a congestion indication (CI) bit. The sources initialize the ER field to their peak cell rate (PCR) and set the CI bit to zero. The switches compute a fair share and reduce the ER field in the returning RM cells to the fair share if necessary. Using exponential weighted averaging a mean allowed cell rate (MACR) is computed and the fair share is set at a fraction of this average:

The problem was fixed by changing to queue growth rate as the load indicator. The change in the queue length is noted down after processing, say, K cells. The overload is indicated if the queue length increases [ 44, 43].

8.3 OSU Time-based Congestion Avoidance

8.4 Congestion Avoidance using Proportional Control (CAPC)

During underload (z<1), fair share is increased as follows:

8.5 Virtual Source and Destination

Figure 7: Virtual source/destination.

Segmenting the network using virtual source/destination reduces the size of the feedback loops. Also, the intermediate segments can use any proprietary congestion control scheme. This allows public telecommunications carriers to follow the standard interface only at entry/exit switches. More importantly, virtual source/destination provide a more robust interface to a public network in the sense that the resources inside the network do not have to rely on user compliance. Misbehaving users will be isolated in the first control loop. The users here include private networks with switches that may or may not be compliant. Notice that the virtual sources and destinations need to maintain per-VC queueing and may, therefore, be quite expensive. There is no limit on the number of segments that can be created. In the extreme case, every switch could act as a virtual source/destination and one would get ``hop-by-hop" rate control as shown in Figure 8.

Figure 8: Hop-by-hop rate control.

8.6 Multicast VCs

Figure 9: Explicit rate control for multicast VCs

9 Credit vs Rate Debate

1. Per-VC Queueing: Credit-based approach requires switches to keep a separate queue for each VC (or VP in case of virtual path scheduling). This applies even to inactive VCs. Per-VC queueing makes switch complexity proportional to the number of VCs. The approach was considered not scalable to a large number of VCs. Given that some large switches will support millions of VCs, this would cause considerable complexity in the switches. This was the single biggest objection to the credit-based approach and the main reason for it not being adopted. Rate-based approach does not require per-VC queueing. It can work with or without per-VC queueing. The choice is left to the implementers.
2. Zero Cell Loss: Under ideal conditions, the credit-based approach can guarantee zero cell loss regardless of traffic patterns, the number of connections, buffer sizes, number of nodes, range of link bandwidths, and range of propagation and queueing delays. Even under extreme overloads, the queue lengths cannot grow beyond the credits granted. The rate-based approach cannot guarantee cell loss. Under extreme overloads, it is possible for queues to grow large resulting in buffer overflow and cell loss. The rate-based camp considered the loss acceptable arguing that with large buffers, the probability of loss is small. Also, they argued that in reality there is always some loss due to errors and, therefore, the user has to worry about loss even if there is zero congestion loss.
3. Ramp-Up Time: The static credit-based approach allows VCs to ramp up to the full rate very fast. In fact, any free capacity can be used immediately. Some rate-based schemes (and the adaptive credit-based approach) can take several round trip delays to ramp up.
4. Isolation and Misbehaving Users: A side benefit of the per-VC queueing is that misbehaving users cannot disrupt the operation of well-behaving users. However, this is less true for the adaptive scheme than for the static credit scheme. In the adaptive scheme, a misbehaving user can get a higher share of buffers by increasing its rate. Note that isolation is attained by per-VC queueing and not so much by credits. Thus, if required, a rate-based switch can also achieve isolation by implementing per-VC queueing.
5. Buffer Requirements: The buffer requirements for the credit-based schemes were found to be less than those in the rate-based scheme with binary feedback. However, this advantage disappeared when explicit rate schemes were added. In the static credit-based approach, per-VC buffer requirement is proportional to link delay, while in the rate-based approach, total buffer requirement is proportional to the end-to-end delay. The adaptive credit scheme allows lower buffering requirements than the static scheme but at the cost of increased ramp-up time. Note that the queueing delays have to be added in both cases since it delays the feedback and adds to the reaction time.
6. Delay estimate: Setting the congestion control parameters in the credit-based approach requires knowledge of link round trip delay. At least, the link length and speed must be known. This knowledge is not required for rate-based approaches (although it may be helpful).
7. Switch Design Flexibility: The explicit rate schemes provide considerable flexibility to switches in deciding how to allocate their resources. Different switches can use different mechanisms and still interoperate in the same network. For example, some switches can opt for minimizing their queue length, while the others can optimize their throughput, while still others can optimize their profits. On the other hand, the credit-based approach dictated that each switch use per-VC queueing with round-robin service.
8. Switch vs End-System Complexity: The credit-based approach introduces complexity in the switches but may have made the end-system's job a bit simpler. The proponents of credit-based approach argued that their host network interface card (NIC) is much simpler since they do not have to schedule each and every cell. As long as credits are available, the cells can be sent at the peak rate. The proponents of the rate-based approach countered that all NIC cards have to have schedulers for their CBR and VBR traffic and using the same mechanism for ABR does not introduce too much complexity.

10 Summary

Acknowledgement

References

1. A. W. Barnhart, ``Use of the Extended PRCA with Various Switch Mechanisms," AF-TM \footnote{Throughout this section, AF-TM refers to ATM Forum Traffic Management working group contributions. These contributions are available only to Forum member organizations. All our contributions and papers are available on-line at http://www.cse.wustl.edu/$\tilde{~}$jain/} 94-0898, September 1994.
2. A. W. Barnhart, ``Explicit Rate Performance Evaluation," AF-TM 94-0983R1, October 1994.
3. J. Bennett and G. Tom Des Jardins, ``Comments on the July PRCA Rate Control Baseline,'' AF-TM 94-0682, July 1994.
4. P. E. Boyer and D. P. Tranchier, ``A Reservation Principle with Applications to the ATM Traffic Control," Computer Networks and ISDN Systems, Vol. 24, 1992, pp. 321-334.
5. A. Charny, "An Algorithm for Rate Allocation in a Cell-Switching Network with Feedback", MIT TR-601, May 1994.
6. A. Charny, D. Clark, and R. Jain, ``Congestion Control with Explicit Rate Indication," AF-TM 94-0692, July 1994. Updated version published in Proc. ICC'95, June 1995.
7. M. Hluchyj, et al., ``Closed-Loop Rate-Based Traffic Management," AF-TM 94-0211R3, April 1994.
8. M. Hluchyj et al, ``Closed-Loop Rate-Based Traffic Management," AF-TM 94-0438R2, September 1994.
9. D. Hughes and P. Daley, ``Limitations of Credit Based Flow Control," AF-TM 94-0776, September 1994.
10. D. Hughes and P. Daley, ``More ABR Simulation Results,'' AF-TM 94-0777, September 1994.
11. D. Hunt, Shirish Sathaye, and K. Brinkerhoff, ``The Realities of Flow Control for ABR Service," AF-TM 94-0871, September 1994.
12. J. Jaffe, ``Bottleneck Flow Control,'' IEEE Trans. Comm., Vol. COM-29, No. 7, pp. 954-962.
13. R. Jain, D. Chiu, and W. Hawe, ``A Quantitative Measure of Fairness and Discrimination for Resource Allocation in Shared Systems,'' DEC TR-301, September 1984.
14. R. Jain, ``The Art of Computer Systems Performance Analysis," John Wiley & Sons, New York, 1991.
15. R. Jain, ``Myths about Congestion Management in High Speed Networks,'' Internetworking: Research and Experience, Vol 3, 1992, pp. 101-113.
16. R. Jain, ``Congestion Control in Computer Networks: Issues and Trends,'' IEEE Network Magazine, May 1990, pp. 24-30.
17. R. Jain, ``A Delay-Based Approach for Congestion Avoidance in Interconnected Heterogeneous Computer Networks,'' Computer Communications Review, Vol. 19, No. 5, October 1989, pp. 56-71.
18. R. Jain, K. K. Ramakrishnan, and D. M. Chiu, ``Congestion Avoidance in Computer Networks with a Connectionless Network Layer,'' Digital Equipment Corporation, Technical Report, DEC-TR-506, August 1987, 17 pp. Also in C. Partridge, Ed., {\it Innovations in Internetworking}, Artech House, Norwood, MA, 1988, pp. 140-156.
19. R. Jain, S. Kalyanaraman, and R. Viswanathan, ``The OSU Scheme for Congestion Avoidance Using Explicit Rate Indication," AF-TM 94-0883, September 1994.
20. R. Jain, S. Kalyanaraman, and R. Viswanathan, ``The EPRCA+ Scheme" AF-TM 94-0988, October 1994.
21. R. Jain, S. Kalyanaraman, and R. Viswanathan, ``The Transient Performance: EPRCA vs EPRCA++," AF-TM 94-1173, November 1994.
22. R. Jain, S. Kalyanaraman, and R. Viswanathan, ``A Sample Switch Algorithm,'' AF-TM 95-0178R1, February 1995.
23. R. Jain, S. Kalyanaraman, R. Goyal, S. Fahmy, F. Lu, ``ERICA+: Extensions to the ERICA Switch Algorithm,'' AF-TM 95-1145R1, October 1995.
24. D. Kataria, ``Comments on Rate-Based Proposal,'' AF-TM 94-0384, May 1994.
25. H. T. Kung, et al, ``Adaptive Credit Allocation for Flow-Controlled VCs," AF-TM 94-0282, March 1994
26. H. T. Kung, et al, ``Flow Controlled Virtual Connections Proposal for ATM Traffic Management," AF-TM 94-0632R2, September 1994.
27. B. Lyles and A. Lin, ``Definition and Preliminary Simulation f a Rate-based Congestion Control Mechanism with Explicit Feedback of Bottleneck Rates," AF-TM 94-0708, July 1994.
28. P. Newman and G. Marshall, ``BECN Congestion Control," AF-TM 94-789R1, July 1993.
29. P. Newman and G. Marshall, ``Update on BECN Congestion Control," AF-TM 94-855R1, September 1993.
30. P. Newman, ``Traffic Management for ATM Local Area Networks," IEEE Communications Magazine, August 1994, pp. 44-50.
31. K. K. Ramakrishnan, D. M. Chiu, and R. Jain, ``Congestion Avoidance in Computer Networks with a Connectionless Network Layer. Part IV: A Selective Binary Feedback Scheme for General Topologies,'' Digital Equipment Corporation, Technical Report DEC-TR-510, August 1987, 41 pp.
32. K. K. Ramakrishnan, ``Issues with Backward Explicit Congestion Notification based Congestion Control," 93-870, September 1993.
33. K. K. Ramakrishnan and J. Zavgren, ``Preliminary Simulation Results of Hop-by-hop/VC Flow Control and Early Packet Discard," AF-TM 94-0231, March 1994.
34. K. K. Ramakrishnan and P. Newman, ``Credit Where Credit is Due," AF-TM 94-0916, September 1994.
35. A. Romanov, ``A Performance Enhancement for Packetized ABR and VBR+ Data,'' AF-TM 94-0295, March 1994.
36. L. Roberts, ``Enhanced PRCA (Proportional Rate-Control Algorithm)," AF-TM 94-0735R1, August 1994.
37. L. Roberts, ``The Benefits of Rate-Based Flow Control for ABR Service," AF-TM 94-0796, September 1994.
38. L. Roberts et al., ``New Pseudocode for Explicit Rate Plus EFCI Support," AF-TM 94-0974, October 1994.
39. L. Roberts, ``Rate-based Algorithm for Point to Multipoint ABR Service,'' AF-TM 94-0772R1, November 1994.
40. S. Sathaye, ``Draft ATM Forum Traffic Management Specification Version 4.0,'' AF-TM~95-0013, December 18, 1994.
41. J. Scott, et al, ``Link by Link, Per VC Credit Based Flow Control," AF-TM 94-0168, March 1994.
42. W. Stallings, ``ISDN and Broadband ISDN with Frame Relay and ATM,'' Prentice-Hall, 1995, 581 pp.
43. H. Tzeng and K. Siu, ``A Class of Proportional Rate Control Schemes and Simulation Results,'' AF-TM 94-0888, September 1994.
44. K. Siu and H. Tzeng, ``Adaptive Proportional Rate Control for ABR Service in ATM Networks," University of California, Irvine, Technical Report 94-07-01.
45. H. Tzeng and K. Siu, ``Enhanced Credit-Based Congestion Notification (ECCN) Flow Control for ATM Networks," AF-TM 94-0450, May 1994.
46. L. Wojnaroski, ``Baseline text for Traffic Management Sub-Working Group," AF-TM 94-0394R4, September 1994.