|M.Sc Student||Arov Michael|
|Subject||Rate Based Flow Control in ATM Networks|
|Department||Department of Electrical and Computer Engineering||Supervisor||PROF. Nahum Shimkin|
This thesis is concerned with efficient algorithms for rate based flow control in communication networks, and in particular of ABR (Available Bit Rate) traffic in ATM (Asynchronous Transfer Mode) networks.
In the last several years some attempts to use control theory techniques for rate control were made. In this thesis we show that such control methods can provide enhanced performance for several problems of flow control including ABR, TCP/IP and more.
An investigation of the ATM network is carried out. The principles of network protocols and implementation issues are translated into control theory terms, which permit us to state the objectives of the rate control problem in these terms. A model of the rate control problem is suggested, this model incorporates the features and emphasizes the difficulties of the rate control problem such as delay, non-linearity, saturation and more. Later on a discussion on the Hop-by-Hop control technique versus End-to-End control technique is carried out and conclusions are presented.
One of the major difficulties of networking, namely the delay in command and response, is translated to the classical time delay problem, which is solved by using a Smith predictor based design with an appropriate modification. Tuning simple controllers such as PI, achieves good performance while keeping the system simple. Further improvement is achieved when the controller design is based on optimal control theory, which enables optimal disturbance rejection in terms of buffer occupancy. A dynamic buffer reference is proposed to prevent buffer saturation. This scheme is based on measurement of available bandwidth and a novel two-time-scale algorithm. Effects of saturating elements inside the control loop such as bounded command are diminished by adapting the controller parameters.
The proposed controllers are examined by simulation, which demonstrates the performance improvement obtained by the suggested algorithms.