عملکرد برنامه های کاربردی TCP بر روی شبکه های ATM با خدمات ABR و UBR - تجزیه و تحلیل شبیه سازی

This paper aims to present a comparative simulation study of the performance of TCP traffic over ATM networks with UBR and ABR services; to identify weaknesses of currently most promising ATM congestion control schemes; and to indicate the requirements for a fairer, simpler and more robust congestion control mechanism. Variety of congestion control schemes under various network configurations and traffic patterns are simulated for a comprehensive analysis. Simulation results show that Early Packet Discard (EPD) wastes buffer capacity and has no provision for fairness, Intelligent Marking [K.Y. Siu, H.Y. Tzeng, Computer Communication Review 24 (5) (1995) 81–106] does not provide adequate fairness among competitive connections, and ERICA+ [R. Jain, S. Kalyanaraman, Y. Goyal, S. Fahmy, R. Viswanathan, ERICA switch algorithm: a complete description, ATM Forum Contribution 96-1172, 1996] is too restrictive and unnecessarily complicated. The paper also indicates that TCP flow control and ATM congestion control do not cooperate well when buffer overflows and retransmission occurs, necessitating further control mechanisms for improving performance.

Asynchronous Transfer Mode (ATM) has emerged as the most promising technology in supporting future broadband multimedia communication services, and the driving force behind this technology is increasingly coming from data networking applications. Since most data applications cannot predict their own bandwidth requirements, they usually require a service that dynamically shares the available bandwidth among all active users. In ATM networks, Unspecified Bit Rate (UBR) service and Available Bit Rate (ABR) service are designed to support traditional data traffic applications, as opposed to voice or video. With UBR, the network only makes available unused network capacity and no commitment is made to UBR source. No feedback concerning congestion is provided. Both delays and variable losses are possible. The ABR is designed to improve the service provided to bursty sources that would otherwise use UBR. An application using ABR can specify a Peak Cell Rate (PCR) that it uses and a Minimum Cell Rate (MCR) that it requires. The network allocates resources so that all ABR applications receive at least their MCR capacity. Any unused capacity is then shared fairly among all ABR sources. The main practical difference between ABR and UBR is that, in the case of ABR, the network provides congestion information to the sources, enabling them to adjust the sending rate to avoid congestion and to achieve high throughput. For ATM to succeed, it needs to support the huge legacy of existing data applications, which invariably employ Transmission Control Protocol (TCP) as the transport layer protocol over the Internet. It is thus of paramount importance to study the performance of data transport protocols TCP over ATM networks with UBR service and ABR service. For data traffic, whose traffic patterns are often highly bursty and unpredictable, congestion control poses a challenging problem for stable and efficient operations. TCP provides a reliable transfer of data using a window-based flow control and error control algorithm. Since the UBR service does not include flow control, it has to rely on TCP's congestion control mechanism when UBR service is employed to carry TCP traffic. However, this TCP/UBR combination is not very effective in terms of resource usage. The main reason is that a TCP segment is often much longer than the length of an ATM cell, and has to be segmented into multiple cells. Even if only one cell of the segment is lost, the entire segment is considered lost and has to be retransmitted because ATM does not provide the cell retransmission function for error recovery. Early Packet Discard (EPD) is introduced to alleviate this problem. With this approach, the switch is equipped with a function to identify and discard cells belonging to the same upper-layer packet in order to protect other packets from corruption. The mechanism is expected to provide considerable throughput improvements when compared with TCP over plain UBR service. Explicit Rate (ER) feedback mechanism has been proven effective for congestion control for high-speed networks [1], [2], [3] and [4] that it has been adopted for ABR traffic management by the ATM Forum [5]. The function of the mechanism is to control the cell emission rate of each source end system using feedback information from the network. If a switch in the network becomes congested, it sends congestion indication to the source end systems. Each source end system then decreases its cell emission rate to avoid buffer overflow at the congested switch. After congestion is relieved, the transmission rate is increased. When TCP uses the ABR service to carry its data, two distinct control algorithm mechanisms exist: the TCP window-based congestion control mechanism controlling the end-to-end loop between TCP components, and the ATM congestion control mechanism controlling the ATM networks. These two mechanisms are independent from each other. It is expected that some form of cooperation between the two mechanisms would help to control traffic flow more effectively. Several efforts have been made to study the performance of TCP over ATM networks. Some of them studied TCP over UBR, some studied TCP over ABR with a particular congestion control scheme. Romanow and Floyd [6] investigated the performance of TCP connections over ATM networks with UBR service, and compared it to the performance of TCP over packet-based networks. The simulation results showed that the effective throughput of TCP over ATM could be quite low when cells are dropped at the congested ATM switch. The low throughput was due to wasted bandwidth as the congested link transmitted cells from corrupted packets. They investigated two packet discard strategies, where Partial Packet Discard improves performance to some extent and Early Packet Discard prevents fragmentation and restored throughput to maximum levels. Simulation of TCP over ATM with ABR service was presented in [7]. The results suggested that the performance of data applications could be severely degraded when the available bandwidth in the ATM network is drastically varied between a maximum and a minimum. Although the performance can be improved to a certain level by deploying large buffers at the ATM switch. Hasegawa et al. [8] evaluated performance of three methods: TCP over UBR, TCP over EPD and TCP over ABR. The simulation experiments showed that TCP over ABR can outperform the other two methods in terms of fairness and throughput if the control parameters of rate-based congestion control algorithm were chosen carefully. However, the simulation also indicates that if the parameter set of the rate-based congestion control was not appropriately used, the congestion would be recognized at TCP level due to packet drops and TCP unnecessarily throttled its window size. The authors pointed out that some modification of TCP might be necessary for further performance improvement. Because most applications use the congestion control provided by TCP, it is important to determine the benefit of using ABR congestion control at the subnetwork layer. Most previous studies on the TCP traffic over ATM emphasize on particular congestion control schemes, employ particular network configurations and use different simulation tools from each other. It is difficult to draw comparative conclusions related to performance under various conditions. This paper presents a comparative simulation study of the performance of TCP traffic over ATM network with UBR and ABR service. A comprehensive set of simulation experiments are performed, employing various congestion control schemes, network configurations, and traffic patterns. TCP effective throughput, packet delay and fairness are used as performance measures for comparison. The paper is organized as follows. Section 1 introduces the background information and issues under study. Section 2 presents a brief review of various congestion control schemes analysed in the paper. Section 3 presents the simulation model and its parameters. Section 4 presents simulation results and analysis. Section 5 points out strengths and weaknesses of the two most promising schemes. Section 6 concludes the paper by indicating the requirements for a fairer, simpler and more effective congestion control scheme.

In this paper we have presented the performance of TCP applications over ATM networks. Simulation results are obtained for TCP over UBR, UBR plus EDP, ABR and IM, ABR and ERICA and ABR and ERICA+. Two network configurations and two traffic patterns are implemented, resulting in four sets of simulations for testing the network performance under various conditions. The performance evaluation is based on the effective throughput, packet delay, fairness, buffer requirement, etc. Our study results can be summarized as follows: • The overall performance of UBR service or UBR with EPD is poor. Without rate adjustment on traffic sources, queue length at congested switch can be excessive. If the buffer overflows and cells are dropped, TCP congestion controls of many connections may unwisely react at the same time causing further performance degradation. Increase in throughput can be achieved with larger switch buffer size for UBR service at the cost of higher packet delay. • Even when the switch buffer is large enough for zero cell loss, the unfairness problem remains severe and is unlikely to be solved under UBR or UBR plus EDP. • Effective throughput for TCP over ATM is improved in EPD since the switch drops entire packets prior to buffer overflow, which prevents the congested link from transmitting useless cells and reduces the total number of corrupted packets. However, the setting of the EPD threshold should be careful to avoid wasting bandwidth in transmitting corrupted packets as well as effectively utilize the switch buffer. • Performance of ABR service with Intelligent Marking, ERICA or ERICA+ is significantly improved comparing with UBR service. The network congestion and maximum queue length is well under control with the employment of the above schemes. Buffer overflow can be avoided with requirement of rather small switch buffer size and high throughput can be achieved. • ERICA+ uses queue length as a metric for indication of the network condition, in addition to the input rate. The network resources are more efficiently utilized and effective throughput is better than ERICA. • Intelligent Marking and ERICA+ achieve similar network throughput. However, the setting of many parameters as well as choosing of many options make ERICA/ERICA+ rather complicated to users. Any inappropriate setting will cause the network performance being degraded to some extent. Intelligent Marking is designed in a much simpler way comparing with ERICA/ERICA+. Simple and clear actions are taken by the switch on a RM cell interval basis. Only a few parameters are kept and a small number of computations are needed at switch. • When the ACR of a VC is limited by its PCR or is throttled by an earlier switch along its path, the VC cannot use up its fairshare, IM attempts to allocate the unused bandwidth to other VCs but IM is not able to do it properly. ERICA+ performs better regarding this aspect since it takes into account individual VC share. • The TCP window based flow control on packets and the ATM rate based congestion control on cells do not cooperate well when buffer overflows and cells are dropped. Fragmentation is the main problem caused by the mismatch when TCP runs over ATM. Additional mechanisms to smooth the mismatch between TCP level and ATM level is necessary. In summary, ERICA+ performs best among the congestion control mechanisms considered, however, it is complex with many parameters. Its interval-based approach increases the complexity of the implementation and can become expensive. Intelligent Marking is simple in that it does not require the switch to keep parameters per VC, however, it has problems with fairness and efficiency. For a best congestion control, it is desirable to adopt a per-queue (rather than per VC) accounting like IM to make the mechanism simple and inexpensive. The cause of the unfairness lies mainly in the wide fluctuations of the allocated ACR. Some mechanism should be introduced to reduce such a variation. Furthermore, congestion is a continuous phenomenon, to have an effective control the control must be smooth and predictive. The hard line separating congestion condition from noncongestion condition is arbitrary and will not help to improve throughput. We have recently proposed Fair Intelligent Congestion Control (FICC) based on our analysis presented in this paper. Following considerations are taken into account in the design of FICC [12] and [13]. Traditionally in most congestion control schemes, different rate allocation policies are employed for congestion period and noncongestion period. Greatly different allocation rates are resulted. When a connection experienced congestion its rate is restricted by the bottleneck switch, while other connections can increase their rate usually to a peak rate. Since the sources that traverse more lines are more likely to experience congested switches, they will be allocated much lower rate than others traversing less links and unfairness is introduced. We believe that it is a weakness to use a fixed queue threshold to arbitrarily divide a network into congestion and noncongestion states. The Fair Intelligent Congestion Control treats rate allocation mechanisms for noncongestion and for congestion period in a similar and consistent manner. If fact it does not have a hard line separating noncongestion from congestion. Instead, it aims for a target operating point where the switch queue length is at an optimum level for good throughput and low delay, and where the rate allocation is optimum for each connection. The rate allocation appropriately reflects the distinction among connections that experience switches at different congestion levels. In order to estimate the current traffic generation rate of the network and allocate it among connections fairly, a Mean Allowed Cell Rate (MACR) per output queue is kept at the switch. An explicit rate is then calculated as a function of the Mean Allowed Cell Rate. The function employed is a queue control function that encourages traffic sources if the target operating point is not reached and discourages the sources if the switch operates beyond its target operating point. The FICC is demonstrated to outperform existing schemes mentioned above with the key advantages of well guaranteed fairness, minimized and bounded buffer requirement, minimized packet delay variation, easy parameter settings, robust to parameter mistuning, and simple implementation.