| dc.description.abstract | In the Internet, the majority of the traffic is elastic, and is characterized by the fact that such traffic can withstand loose bounds on performance guarantees. It is a well-understood fact, however, that once the performance drops below a minimum acceptable quality, users of the elastic applications do not derive utility from them. This is especially true when the number of elastic sessions sharing a scarce resource increases to a large value. Unless the bandwidth is properly managed, the resource will be used inefficiently, and the users will experience poor Quality of Service (QoS).
Users of elastic applications are generally not sensitive to the end-to-end delay of each packet, but to the time necessary to transfer an entire file, which critically depends on the number of such flows sharing the network bandwidth. One way to achieve a predefined notion of QoS is through admission control, which consists of rejecting new flows during overload conditions. In this thesis, we propose a few adaptive algorithms to implement admission control.
We begin by studying bandwidth sharing of elastic flows at a bottleneck link. For this, we model the bottleneck link by an M/G/1 PS queue and consider a particular notion of throughput measure and study it in the above scenario. We then consider blocking new connections to implement admission control. Here, blocking is implemented as a probability. Towards this end, we model online occupancy measurements of the link and use these measurements to estimate the offered traffic at the link, which in turn is used to compute the blocking probability value. In this category, we propose two algorithms, namely, least squares estimate and exponential averaging, to estimate the traffic at the bottleneck link.
Generally, when a new connection is rejected as a part of the admission control, it is not necessarily lost but retries after some time. Thus, retrials constitute another arrival process apart from the fresh arrival process. In such a case, the controlled traffic is related to the blocking probability value through a nonlinear relation. For this problem, we propose a few stochastic approximation-based algorithms and apply them to the traffic control problem.
For all the algorithms proposed in the thesis, we show their efficacy by the results obtained from simulation experiments. | |
