ABSTRACT

The broadband integrated services networks will probably be completed through the realization of the global public ATM network, which is to be used for the Ethernet LAN interconnections. In the heterogeneous networks of such structure the conditions for the data traffic growth are to be achieved, where the flow control algorithms take the main role in the process of integration. The ATM network flow control, realized through the ABR (with feedback and QoS guarantees) and UBR (without feedback and QoS guarantees) class of services, takes place on the homogeneous part of the network. When the parameters of the flow control are known on the edge of the ATM network (ABR), they cannot be signaled to the end stations using the existing protocols, resulting in the sending rate adaptation difficulties. The solutions are found in the modified flow control architecture in three variants: without parameter signaling, but with data shaping on the edge of the ATM network, with parameter signaling and flow control criteria unification on the heterogeneous networks, and with local feedback loop closing on the homogeneous parts of the network The first variant, providing simplicity and independence of the network type and being the only one suitable for the UBR class of service, has been given preference to.

The optimal flow control criteria are satisfactory network utilization and minimal user data (packets) delay. Because of the stochastic character of the traffic generation process, as well as of the servicing process in the network, the possibility of packet losses, due to the node queues overload increases with the network utilization. Using the G/G/1 queuing model and uniform distribution instead of general (with mean and variance preserved), numeric results have been obtained for the delay curves with relative variance as parameter. Two traffic management strategies have been developed: the maximum power and the constant QoS strategy. In future networks, the user traffic should be shaped to keep the variance of arrival and service times low enough, so that a high network utilization can be obtained. In the ideal situation, the network can be modeled with an D/D/1 queuing system.

In this work, taking into account the above conclusions, the network has been modeled with the ideal D/D/1 queuing system. The black box approach has been used, being universal in view of the heterogeneous network model. The D/D/1 system has been adapted with total number of packets in the network limitation (D/D/1/W), which has made possible the introduction of the WT (Window-Time) space. The delay curves of the D/D/1 system are transferred from the rT (Load-Time) to the WT space. The analytical expressions for the D/D/1/W model in the WT space have been developed. It has been assumed that the user would utilize only a part of the total network (path) capacity. The mathematical model of the network response for the capacity lower than the maximal has been developed, and a family of curves depending on the partial path capacity has been obtained.

It has been shown that, measuring a point coordinates in the WT space, while assuming the information on the total network (path) capacity to be available, the sender can unambiguously estimate the optimal packet sending rate and window, using these parameters for the optimal usage of the available network capacity. The analytical expressions for optimal rate and window estimates under the network underload and overload conditions have been developed. The precision of the available capacity estimate depends on the knowledge about the total network (path) capacity. Three algorithms of the total capacity correction have been proposed, and their stability has been analyzed.

Using the theoretical results described above the optimal flow control algorithm for the D/D/1/W network model has been proposed. Besides directly using basic scientific developments additional original algorithms have been proposed, such as packet sending at start-up and sending rate filtering. This has made possible to obtain a formal specification of the consistent flow control algorithm using the pseudo programming language.

The proposed optimal flow control algorithm is embedded in the packet network simulator NS1.1 from Lawrence Berkeley Laboratory, CA, USA. The simulation tests have been performed systematically using four basic network topologies with 1, 2 or 3 user connections, with channel capacity variations from 15 kb/s to 15 Mb/s and channel delay variations from 10 ms to 1000 ms.

The simulation measurement analysis has shown that the proposed flow control algorithm efficiently and optimally uses the available network capacity on the conventional as well as on the dynamically changing (ATM) channels. During the operation it sends to the network exactly the optimal window of packets, thus minimizing the waiting times in the queues. When the second and third connection were activated, and the third cleared subsequently, all users detected a change in the available capacity and efficiently adapted their sending rate and window.

The scientific contribution in this dissertation has been obtained through the analysis of the existing network models and flow control methods in information networks, using it to define the optimality criteria for the flow control algorithm development. The algorithms should ensure the communicating process sender operation with as low the packet arrival period variance as possible. During the research the basic scientific contribution has been obtained through the definition of the WT space congestion network model, the available network capacity estimation, the achievement of the optimal sending rate and window and the estimation of the total network (path) capacity. The optimal flow control algorithm has been proposed using the theoretical model, as well as the original mechanisms for start-up, window correction and rate filtering. The analysis of the simulation measurement results of the proposed algorithm has been made, and the conclusions on its functionality have been developed.