Optical burst switching (OBS) is considered as a contending technology for the core of the Internet in future. However, due to lack of the buffers, losses occur due to contention among simultaneously arriving bursts at the core nodes. Contention losses do not necessarily indicate a situation of congestion in the network. Thus differentiation (classification) of losses is essential in many applications to avoid false identification of congestion. In this paper, we propose a loss classification technique for the OBS networks based on machine learning techniques. We devise a new measure to differentiate between congestion and contention losses, which is derived from the observed losses, called the number of bursts between failures (NBBF). We observe that the NBBF follows a Gaussian distribution with different parameters for contention and congestion losses. This feature is used in differentiation. We use both a supervised learning technique (hidden Markov model (HMM)) and an unsupervised learning technique (expectation maximization (EM) clustering) on the observed losses and classify them into a set of states (clusters) after which an algorithm differentiates between the congestion and contention losses. We also demonstrate the use of loss differentiation in improving the performance of transport control protocol (TCP) over OBS networks. We modify congestion control mechanism of TCP suitably to arrive at two variants of TCP, HMM-TCP and EM-TCP. Their performance is compared with TCP NewReno, TCP SACK, and Burst TCP (X. Yu et al., Mar. 2004). Simulation results demonstrate the effectiveness and accuracy of the loss classification technique in different network scenarios.

It is widely accepted that retransmitting bursts in Optical Burst Switching networks improves the throughput at higher layers. Since the data is transported in large bursts and there are burst losses due to contention, indiscriminate retransmission can defeat the purpose of burst-level retransmissions. For many applications, retransmission after a certain time would be of no use and it may not be necessary to retransmit all the lost packets. We propose a framework for retransmission with parameters to control the retransmission rate and thus the increase in the network load. We propose a network model for controlled retransmission and a modified functional architecture of the ingress node. The existing work lacks an accurate analysis to estimate the impact of retransmissions on the network load while studying the improvement in the end-to-end packet recovery. We provide theoretical analysis to evaluate the load at each node due to both fresh and retransmitted bursts. We propose some metrics to quantify the benefit of retransmission and the impact of the retransmission parameters proposed on the network performance. We show that with controlled retransmission, the buffer requirement at the ingress node is proportional to the steady-state rate of retransmitted bursts. We validate the analytical model arid also study the impact of retransmission on the network performance with extensive simulations.