Reactive traffic control mechanisms for communication networks with self-similar bandwidth demands
Thesis DisciplineElectrical Engineering
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
Communication network architectures are in the process of being redesigned so that many different services are integrated within the same network. Due to this integration, traffic management algorithms need to balance the requirements of the traffic which the algorithms are directly controlling with Quality of Service (QoS) requirements of other classes of traffic which will be encountered in the network. Of particular interest is one class of traffic, termed elastic traffic, that responds to dynamic feedback from the network regarding the amount of available resources within the network. Examples of this type of traffic include the Available Bit Rate (ABR) service in Asynchronous Transfer Mode (ATM) networks and connections using Transmission Control Protocol (TCP) in the Internet. Both examples aim to utilise available bandwidth within a network. Reactive traffic management, like that which occurs in the ABR service and TCP, depends explicitly on the dynamic bandwidth requirements of other traffic which is currently using the network. In particular, there is significant evidence that a wide range of network traffic, including Ethernet, World Wide Web, Varible Bit Rate video and signalling traffic, is self-similar. The term self-similar refers to the particular characteristic of network traffic to remain bursty over a wide range of time scales. A closely associated characteristic of self-similar traffic is its long-range dependence (LRD), which refers to the significant correlations that occur with the traffic. By utilising these correlations, greater predictability of network traffic can be achieved, and hence the performance of reactive traffic management algorithms can be enhanced. A predictive rate control algorithm, called PERC (Predictive Explicit Rate Control), is proposed in this thesis which is targeted to the ABR service in ATM networks. By incorporating the LRD stochastic structure of background traffic, measurements of the bandwidth requirements of background traffic, and the delay associated with a particular ABR connection, a predictive algorithm is defined which provides explicit rate information that is conveyed to ABR sources. An enhancement to PERC is also described. This algorithm, called PERC+, uses previous control information to correct prediction errors that occur for connections with larger round-trip delay. These algorithms have been extensively analysed with regards to their network performance, and simulation results show that queue lengths and cell loss rates are significantly reduced when these algorithms are deployed. An adaptive version of PERC has also been developed using real-time parameter estimates of self-similar traffic. This has excellent performance compared with standard ABR rate control algorithms such as ERICA. Since PERC and its enhancement PERC+ have explicitly utilised the index of self-similarity, known as the Hurst parameter, the sensitivity of these algorithms to this parameter can be determined analytically. Research work described in this thesis shows that the algorithms have an asymmetric sensitivity to the Hurst parameter, with significant sensitivity in the region where the parameter is underestimated as being close to 0.5. Simulation results reveal the same bias in the performance of the algorithm with regards to the Hurst parameter. In contrast, PERC is insensitive to estimates of the mean, using the sample mean estimator, and estimates of the traffic variance, which is due to the algorithm primarily utilising the correlation structure of the traffic to predict future bandwidth requirements. Sensitivity analysis falls into the area of investigative research, but it naturally leads to the area of robust control, where algorithms are designed so that uncertainty in traffic parameter estimation or modelling can be accommodated. An alternative robust design approach, to the standard maximum entropy approach, is proposed in this thesis that uses the maximum likelihood function to develop the predictive rate controller. The likelihood function defines the proximity of a specific traffic model to the traffic data, and hence gives a measure of the performance of a chosen model. Maximising the likelihood function leads to optimising robust performance, and it is shown, through simulations, that the system performance is close to the optimal performance as compared with maximising the spectral entropy. There is still debate regarding the influence of LRD on network performance. This thesis also considers the question of the influence of LRD on traffic predictability, and demonstrates that predictive rate control algorithms that only use short-term correlations have close performance to algorithms that utilise long-term correlations. It is noted that predictors based on LRD still out-perform ones which use short-term correlations, but that there is Potential simplification in the design of predictors, since traffic predictability can be achieved using short-term correlations. This thesis forms a substantial contribution to the understanding of control in the case where self-similar processes form part of the overall system. Rather than doggedly pursuing self-similar control, a broader view has been taken where the performance of algorithms have been considered from a number of perspectives. A number of different research avenues lead on from this work, and these are outlined.