IMA: پایه های فنی، نرم افزار و تجزیه و تحلیل عملکرد

Using WAN established infrastructure, one of the main problems ATM network planners and users face, when greater than T1/E1 bandwidth is required, is the disproportionate cost associated with T3/E3 links. The technology to cover the gap between T1/E1 and T3/E3 bandwidth at a reasonable cost is known as inverse multiplexing for ATM (IMA). IMA allows multiple T1/E1 lines to be aggregated to support the transparent transmission of ATM cells over one single virtual trunk. In this paper, the fundamentals and major applications of IMA technology are described. Also, the behavior of IMA multiplexers is carefully analyzed and a method to dimension them proposed. For this purpose, an IMA simulation tool has been developed. The IMA simulator permits the study of individual devices and the evaluation of the end-to-end performance of a logical trunk under several ATM input traffic patterns. The analytical study is based on the comparison with an M/D/C/(N+C) queue system. Under Poisson input traffic, an approximation for the cell loss ratio (CLR) is derived and an estimate of the cell delay in an IMA multiplexer obtained. In addition, the suitability of these results for two types of bursty traffic is investigated.

The growing demand for high speed services is accelerating broadband integrated services digital network (B-ISDN) deployment to support conventional and new data, voice, video and multimedia applications in a single network. This network is based on the asynchronous transfer mode (ATM) to carry any type of traffic efficiently. The dominant role of ATM in the LAN scenario is still unclear in comparison with other competing technologies such as switched Ethernet, Fast Ethernet and even Gigabit Ethernet. ATM is well introduced in the backbone area, especially in private corporate environments. Regarding WAN, network operators are first providing access to the ATM network using the existing infrastructure and then gradually deploying the new one, implementing pure ATM interfaces over it. Users, however, want the ATM bandwidth benefits for their high speed applications soon, but in a cost-effective manner. Currently, there are basically two available options to provide access to the ATM services on a WAN scale. One consists of T3/E3 links offering considerable bandwidth (44.736/34.368 Mbps) but is usually not justified since it would be underutilized by most of the prospective users. Furthermore, the rates that carriers charge for them are very high. The other alternative is substantially cheaper and uses T1/E1 links (1.544/2.048 Mbps), but the offered bandwidth is insufficient for some user needs. Prices depend on several factors such as distance and each particular carrier. As an example, the average cost per month of a 25 km T1/E1 link is $850/$2900, respectively, and $7500/$29,000 for a T3/E3 line of the same length [7]. However, in general, T3/E3 links have their point of presence and are only available in big cities [7] and [14]. Due to cost and availability of service, an intermediate solution offering enough bandwidth at a reasonable cost is required. In July 1997, the ATM Forum published the inverse multiplexing for ATM specification, known as IMA [3], the last version of which was released in April 1999 [4]. IMA defines the transparent transmission of a high speed ATM cell stream over one logical link composed of several T1/E1 lines. IMA distributes and transfers a single flow of ATM layer cell traffic onto multiple physical links. At the remote end, the traffic is recombined and the original ATM cell sequence fully recovered and delivered to the higher layers that will further process it. Up to 32 T1 or E1 links can be used to form an IMA group that operates at an aggregated bit rate of some multiple of the T1/E1 speed. Up to 48/64 Mbps can be reached. These bit rates are enough to support many current user broadband applications requiring a fractional T3/E3 bit rate but using bandwidth more efficiently, and utilizing readily available and less expensive T1/E1 services. The inverse multiplexer (IMUX) is the device responsible for grouping several T1/E1 physical circuits into a single logical trunk. An IMUX accepts ATM cell streams coming from different traffic sources, in addition to traffic coming directly from LANs (e.g., from a router without an ATM interface). This non-ATM traffic is adapted and converted to ATM cell format, using ATM layer segmentation and re-assembly functionality. In both cases, the IMUX distributes the resulting cells in round-robin fashion over the physical links maintaining the QoS required by each individual connection. To configure, control, maintain and synchronize the links belonging to an IMA group, the IMUX introduces two types of operation and maintenance (OAM) cells. That is, IMA control protocol (ICP) and Filler cells. Thus, in this paper, the origins, application and technical foundations of inverse multiplexing are explained in tutorial style. Then, a model for an IMUX is presented and intensively evaluated. To perform the evaluation under different input traffic distributions an IMA system simulator was developed. The idea was to elaborate a methodology to help engineers and network planners to characterize and dimension an IMUX device, that is, to obtain the buffer size and the number of T1/E1 output links that guarantee the required QoS parameters demanded by users, basically measured as cell loss ratio (CLR) and average cell delay. An approximate analysis allowing easy computation of the IMUX performance was derived, obviating the need to perform costly simulations. The IMUX dimensioning study was conducted under Poisson input traffic. More realistic traffic patterns are also presented in this study. Due to their simplicity, we decided to use two models of bursty traffic instead of those described in other surveys [5], [9], [19] and [20]. These bursty models are an on–off pattern [6], and a WWW traffic characterization [11] for residential networks. Of course, the usefulness of these patterns is limited to certain scenarios (e.g., Poisson traffic is a characterization of multiple traffic aggregations), but equally obviously, the performance of the IMUX is traffic dependent. For this reason and since the actual traffic behavior is subject to short-term future changes depending on user requirements, we present this study as a more “permanent” way to model IMUX performance. This research attempts to help engineers to validate and verify their device node libraries (IMUX in this case) in the network simulators they have to develop, since no other means (e.g., mathematical analysis) is now available in a closed, useful and easy manner to plan broadband networks in a traffic changing environment. Although several relevant traffic characterization advances have been achieved, they show well-known characteristics (e.g., its self-similar nature), and the resulting models are too complicated to help in the analysis of emerging devices and networks. Once these libraries are validated, engineers can evaluate IMUX performance with more realistic, but changing, traffic patterns. The paper is organized as follows. Section 2 discusses the current knowledge regarding IMUX operation and its main applications. The simulation model is introduced in Section 3. The study under Poisson input traffic is presented in Section 4. This section also includes the approximate analysis for the CLR and the average cell delay. The results for bursty traffic are described in Section 5. Finally, Section 6 summarizes the main conclusions of this study and outlines some future work.

In this paper, inverse multiplexing for ATM and its primary network applications are described. IMA provides an economical alternative to the bandwidth gap between T1/E1 and T3/E3. Using IMA, two or more T1/E1 can be combined to create a single logical connection with an effective bandwidth of multiple T1/E1 lines. ATM cells are multiplexed and de-multiplexed in round-robin fashion among an IMA group of output links. IMA represents a physical layer technology; therefore, it can be used to transport any service once it has been adapted to ATM cell format. To fulfil its function in both directions of communication, IMA builds IMA frames and inserts control cells (ICP, SICP and Filler cells). These allow the introduction of time markers to recover the original ATM layer cell stream, cell decoupling rate function and synchronism operations. To study the behavior and evaluate the performance of IMA systems a flexible and object-oriented simulation tool has been developed. IMA multiplexers have been characterized and analyzed under Poisson input traffic. First, the CLR as a function of the offered load and number of output links is investigated. An equivalent M/D/C/(N+C) queue model is used to propose a method to dimension the memory required at the IMUX. Reducing the queue size of this analytical model, an accurate approximation to the simulation results is obtained. The queuing system is further manipulated to adjust the average cell delay. The fit achieved is good enough, although the method is dependent on the simulation results. We are currently working towards mathematical approximations and exact solutions of M/D/1/N queue systems and obtaining very promising results, regarding both CLR and average cell delay. In these cases, the need to conduct costly simulations is completely obviated. Preliminary results for two types of bursty traffic patterns have also been presented. These traffic models are characterized by the generation of cells in active periods and inactivity during idle periods. In general, more buffering is needed to cope with this traffic. In some cases, the amount of memory is prohibitive and performance does not improve on increasing it. Under on–off traffic and for short and medium bursts the proposed methodology to dimension IMUX resources is still valid. However, when the injected traffic is a characterization of WWW services over residential networks the amount of memory required at the IMUX for a given QoS is excessive. This is due to long active periods compared to the time slot in the output lines, even for low applied loads (<1%). Nevertheless, IMUX performance depends on the input traffic pattern. This also holds for any network node. Using Poisson traffic is almost the only way to obtain analytical models and closed mathematical expressions. Although Poisson traffic is not a realistic pattern for access network nodes and for the current traffic measured in ATM networks, it is still a useful approach to validate the characterization and simulation models of these ATM network nodes. Only a few years ago, no one forecasted the current traffic patterns on the Internet, mainly WWW traffic. Therefore, the actual traffic behavior is subject to near future changes in user behavior, resulting in self-similar patterns or others. However, operators, network planners and traffic engineers will continue developing device node libraries to incorporate more or less complex network simulators. These simulators will be fed with realistic traffic patterns according to users’ changing needs. However, it is highly probable that engineers will still use easy and useful patterns, such as the Poisson one, to validate and verify their newly developed simulators. This is why we still believe in the usefulness of a “permanent” and understandable traffic pattern to study the IMUX and to analyze it mathematically [1] and [18]. In this paper, we show the changing behavior of our ATM IMUX node for other input traffic for the sake of being complete. It can be concluded that the dimensioning procedure proposed is reliable enough for short and medium length active periods. At present, we are examining the actual degree of correctness and the extrapolation of results under different bursty traffic patterns, together with possible solutions when the method does not hold. The new patterns to investigate consider single and multiplexed traffic sources. Simple but realistic traffic patterns that are understandable by network planners are especially relevant to our research. Finally, another part of our current work is the incorporation of different traffic classes into the simulation environment (CBR, VBR-rt, VBR-nrt, ABR and UBR). This will involve some changes in the buffering schemes used and the addition of scheduling mechanisms to handle ATM cells. In this new and more realistic framework, we will extend our analysis to approximate the performance offered by IMA systems.