this paper, we will assume that all window sizes are measured in units of maximum size packets, instead of bytes. In the original TCP specification [10], the window used by the sender, which we will denote by wnd, is the 1

Performance evaluation of TCP over ATM has been the focus of much research recently. A number of studies have analysed observed performance degradation, suggested ways to improve throughput, compared the performance of ABR and UBR as suitable class of ATM service in transporting TCP traffic. With this explosion of studies exploring various combinations of scenarios (LANs versus WANs, backbone ATM versus native ATM to mention few), there is a need for a framework for a coherent formulation of the design space and analysis of the results. The aim of this paper is to characterise the major research issues and draw upon the published work to illustrate some of the proposed solutions. The aim is to move towards a framework for building a taxonomy, and this paper is the first step towards achieving this. The characterisation of the performance evaluation of TCP over ATM is undertaken by identifying and evaluating the research problems, the experimental designs to address the prob...

this paper is to understand the results in Reference [18], which we reviewed briefly in Section 3.2. Are those results explained by what we have seen in our simple two-way traffic configurations? Compare Figure 3 with Figure 6. The rapid queue fluctuations in Figure 3 are similar to those in Figure 7, indicating the presence of ACK-compression. Furthermore, the synchronization and idle time apparent in Figure 3 resembles those of the out-of-phase synchronization mode in Figure 6. These were the key features we wanted to understand. There are some differences, however, between the data in Figure 3 and that in Figure 6; in Figure 3 the plateaus of the square-wave-like fluctuations are narrower, the queue length rise more rapid, and the dynamics significantly less regular. The widths of the plateaus reflect the sizes of the clusters. Recall that the configuration analyzed in Figure 3 had a buffer of size 30, with five connections in each direction and ø = 0:01 sec. Thus, if the dynamics were completely regular and symmetric, each connection would have wnd ß 6 during the congestion epoch. This is in contrast to wnd values of 20 or more for the simpler configurations considered in this paper. This explains the narrowness of the plateaus. The rate at which the queue length rises is related to the total acceleration and the total acceleration during a congestion epoch is just the total number of connections. Since we have 10 connections in the configuration for Figure 3 compared to just 2 for Figure 6, we would expect the queue length to rise much more rapidly in Figure 3. The regularity in the simple configurations considered in this paper is due to the complete clustering of the packets. We have explained in Reference [15] why this clustering occurs for one-way traffic confi...