Performance of TCP over UBR in ATM with EPD and virtual queuing techniques

Abstract

While both available bit rate (ABR) and unspecified bit rate (UBR) services can be used to support data traffic in asynchronous transfer mode (ATM), many ATM switch vendors consider UBR to be more attractive because of its low implementation cost in comparison with ABR. However, since there is no congestion control at the cell level in UBR and cells are simply discarded when buffer overflow occurs, the effective throughput of transmission control protocol (TCP) can degrade significantly over UBR service in a congested network. Early Packet Discard (EPD) techniques have been proposed by Romanow and Floyd (Dynamics of TCP traffic over ATM networks, IEEE Journal on Selected Areas in Communications 13 (4) (1995) 633–641) and shown to improve the throughput of TCP over ATM. In our earlier work (A simulation study of TCP performance in ATM networks with ABR and UBR services, in: Proceedings of the International Phoenix Conference on Computers and Communications, March 1996), we have shown that TCP with EPD can suffer significant degradation in fairness of throughputs among competing VCs in a congested ATM LAN environment, but the degree of fairness can be improved by using per-VC accounting/queuing techniques. In this paper, we continue our studies in designing better and low cost methods for the support of TCP over ATM. We apply the virtual queuing technique proposed in Chiussi et al. (Virtual queuing techniques for ABR service: improving ABR/VBR interaction, Infocom'97) to emulate on a FIFO queue the service provided by per-VC queuing. This technique requires low implementation cost and delivers performance of TCP over UBR comparable to the expensive per-VC queuing technique. Simulation results are presented to demonstrate that our technique combined with EPD can drastically improve the performance of TCP over UBR service.

Introduction

While asynchronous transfer mode (ATM) was originally conceived as a carrier of integrated traffic, the recent momentum on the rapid standardization of the technology has come from data networking applications. Since most data applications cannot predict their own bandwidth requirements, they usually require a service that allows all competing active virtual connections (VCs) to dynamically share the available bandwidth. Unspecified bit rate (UBR) and available bit rate (ABR) are two types of services standardized for the support of data applications in ATM networks. Since transmission control protocol (TCP) is perhaps the most widely used transport layer protocol in existing data networks, the performance of TCP over ABR and UBR services in ATM is of major interest to ATM equipment vendors and service providers.

UBR service is designed for those data applications that want to use any available bandwidth and are not sensitive to cell loss or delay. Such connections are not rejected on the basis of bandwidth shortage and not policed for their usage behavior. During congestion, the cells may be lost but the sources are not expected to reduce their cell rates. Instead, these applications may have their own higher-level loss recovery and retransmission mechanisms, such as the window flow control employed by TCP.

Although it is relatively simple to support UBR service in ATM switches, recent results suggest that UBR without any ATM layer congestion control mechanism does not yield adequate performance [1]. The reason for such performance degradation is that when the loss mechanism simply discards arriving cells when buffer overflow occurs, each discarded cell is likely to belong to a different packet. A significant portion of the available bandwidth is then wasted since it is occupied by cells belonging to packets that are already corrupted by cell loss and thus need to be retransmitted.

To address this problem, more sophisticated frame discarding mechanisms such as the Early Packet Discard (EPD) algorithm [1]for UBR service have been proposed. The basic idea of EPD is to discard an entire packet prior to buffer overflows, so that the bandwidth is utilized for the transmission of only non-corrupted packets. EPD in fact represents a general class of packet discarding algorithms, which can be applied to any packet-based protocol running over ATM, such as TCP, UDP, and IPX, etc.

One drawback of the EPD scheme proposed in Ref. [1]is that it cannot achieve fair bandwidth allocation among competing VCs [2]. For example, EPD tends to allocate less bandwidth for VCs with longer round-trip time and for VCs traversing multiple congested nodes. This is often referred to as the `VC starvation' problem. In addition, connections with bulky data tend to get more bandwidth than highly bursty traffic connections under EPD.

In our earlier work [3], we presented a study on extending EPD by incorporating a fair buffer allocation mechanism to alleviate the fairness problem. Two fair buffer allocation techniques have been studied: per-VC accounting and per-VC queuing techniques. Per-VC accounting uses a single FIFO queue and state information on a per-VC basis. The per-VC state information can include, for example, the number of cells of each VC in the FIFO queue. On the other hand, per-VC queuing implements a separate queue for each VC and permits output cell scheduling on a per-VC basis such as round-robin cell scheduling. In [3]we presented simulation results on the performance of TCP over UBR service using EPD combined with per-VC accounting and per-VC queuing techniques. Compared with the simple EPD mechanism, significant improvement of performance, particularly in terms of fairness, using these fair buffer allocation techniques has been observed. From the simulation results, we can also conclude that per-VC queuing indeed offers far better performance than the specific per-VC accounting technique examined in Li et al. [3]. However, per-VC queuing is much more expensive to implement than a FIFO queue with per-VC accounting. Thus despite its performance advantage, per-VC queuing is not implemented in many existing ATM switches for UBR service.

In this paper, we apply the virtual queuing technique proposed by Chiussi et al. [4]to emulate per-VC queuing on a FIFO queue. This application of the virtual queuing technique can be viewed as a sophisticated form of per-VC accounting and it uses only a simple data structure. We shall illustrate with simulation results how the technique can be applied to significantly improve the performance of EPD algorithms for UBR service.

Besides UBR service, an alternative approach to support data traffic in ATM is to use ABR service. Whereas EPD for UBR service is an open-loop mechanism, the current standard for ABR service uses a closed-loop rate-based mechanism for congestion control. The general idea of the mechanism is to adjust the input rates of sources on the basis of the current level of congestion along their VCs. A number of different rate-based congestion control schemes have been proposed, and they differ from one another by the various switch mechanisms employed to determine the congestion information and the mechanisms by which the source rates are adjusted.

Many of the promising rate-based proposals 5, 6, 7, 8use a technique called `intelligent marking' originally proposed in 9, 10. The key idea of intelligent marking is to estimate the bottleneck rate of each VC and achieve max–min fairness without the need of per-VC queuing nor per-VC accounting. Subsequent proposals following the work of Siu and Tzeng 9, 10have incorporated per-VC accounting to keep track of the input and output cell rates at a switch on a per-VC basis. Per-VC accounting enhances the accuracy in estimating the bottleneck rate for each VC, at the expense of additional complexity in switch implementation. A survey of these rate-based congestion control techniques can be found in Refs. 7, 11.

The remainder of the paper is organized as follows. In the next section, we shall discuss further related works on packet discard algorithms and TCP over ATM, and results in other areas that inspire our ideas of virtual queuing. The EPD and its enhanced mechanisms, including our virtual queuing technique, are described in Section 3. Our simulation model and parameters are discussed in Section 4. The simulation results comparing the performance of various EPD techniques are presented in Section 5. Concluding remarks are given in Section 6.