Adaptive algorithms for admission control of elastic sessions in the internet

Abstract

Elastic traffic dominates Internet usage and is characterized by its tolerance to loose performance guarantees. However, when network performance drops below a minimum acceptable threshold - especially under high load - users experience poor Quality of Service (QoS), leading to inefficient resource utilization. Unlike delay-sensitive applications, elastic users are more concerned with file transfer completion times, which depend heavily on bandwidth sharing among concurrent flows. This thesis proposes adaptive algorithms for admission control to manage bandwidth and maintain QoS for elastic traffic. The study begins by modeling bandwidth sharing at a bottleneck link using an M/G/1 Processor Sharing (PS) queue, introducing a throughput-based performance measure. Admission control is implemented via probabilistic blocking, where online occupancy measurements are used to estimate offered traffic and compute blocking probabilities. Two estimation techniques - least squares and exponential averaging - are proposed for this purpose. Recognizing that rejected connections often retry, the thesis models retrials as a secondary arrival process, introducing a nonlinear relationship between traffic and blocking probability. To address this, stochastic approximation algorithms are developed and applied to control traffic effectively. Simulation experiments validate the proposed algorithms, demonstrating their effectiveness in maintaining QoS and optimizing bandwidth usage under varying network conditions.