Machine Learning Algorithms for Anomaly Detection in Optical Networks

Tarem Ahmed, Department of Electrical and Computer Engineering, McGill University

Mark Coates, Department of Electrical and Computer Engineering, McGill University

Optical networks are known to experience a wide range of anomalous behaviour and network measurements are seen to be of non-stationary nature.  Machine learning techniques enable the development of anomaly detection algorithms that are non-parametric and adaptive to changes in the characteristics of normal behaviour, making the algorithms robust and portable across applications.  In this research we develop the recursive Kernel-based Online Anomaly Detection (KOAD) algorithm and apply it to anomaly detection in large optical backbone networks.  The KOAD algorithm has been extended from our earlier online anomaly detection algorithm based on Kernel Recursive Least Squares, and shown to produce quick detection with high accuracy and low false alarm rates.  We also investigate relationship between the region of normality identified by our KOAD algorithm and the concept of minimum volume sets, and apply the block-based One-Class Neighbour Machine algorithm developed by Muñoz and Moguerza.  We test our algorithms on a time series of entropies of the main IP packet headers traversing the Abilene backbone network.  The entropy statistic captures the distribution of the traffic, and spotting sudden changes in the entropy enables the detection of a wide range of anomalies.

Minimum Cost Dimensioning of Ring Optical Networks

Alain C. Houle, GEGI, Université de Sherbrooke
Brigitte Jaumard, CIISE, Concordia University

Abdallah Jarray, DIRO, Université de Montréal 

We consider the problem of traffic grooming of low-rate traffic circuits in WDM rings where circuits are associated with a set of heterogeneous granularities. While networks are no longer limited by transmission bandwidth, the key issue in WDM network design has evolved towards the processing capabilities of electronic switches, routers and multiplexers. Therefore, we focus here on traffic grooming with minimum interconnecting equipment cost. We first formulate the problem as an integer linear programming (ILP) or a mixed integer linear programming (MILP) problem depending on the design specifications: UPSR vs BLSR, fixed vs variable wavelength capacities, non bifurcated vs bifurcated flows, wavelength continuity vs possible signal regeneration on a different wavelength. Considering the case study of the second SONET ring generation with MSPP like interconnection equipments, we define the cost by a function of the number of transport blades, taking into account that the number of MSPP transport blades makes up a significant portion of the overall network design cost. Using the CPLEX linear programming package, we next compare the optimal solutions of the ILP or MILP programs for different design assumptions, including the classical ring network design scheme with a single hub where the lightpaths directly connect the hub to all other nodes.

A Control Plane for an Agile All-Photonic Network

Mahdi Hermas, Department of Electrical and Computer Engineering, McGill University

Imad Khazali, Department of Electrical and Computer Engineering, McGill University

Richard Vickers, Department of Electrical and Computer Engineering, McGill University

The Agile All-Photonic Network (AAPN), proposed for the next generation optical transport networks, is an innovative, self-configuring, distributed switch that is based on the latest advances in optical switching technology. AAPN’s architecture and low switching granularity allows it to efficiently support connection-oriented and connectionless traffic between its high speed edge nodes. This work describes the conceptual and functional requirements of a control plane that suits AAPN’s architecture. It considers routing (e.g. neighbour and topology discovery) and signalling for dynamic provisioning and restoration issues. Also, control plane requirements unique for AAPN’s architecture are described.

Extended Reserve-and-Preempt Scheme for Providng Absolute QoS Differentiation on OBS networks

Hongbo Lui, SITE -  School of Information Technology and Engineering,  University of Ottawa 

Hussein Mouftah, SITE -  School of Information Technology and Engineering, University of Ottawa 

Providing absolute QoS differentiation is an important issue in OBS networks. While satisfying the loss thresholds of each of the priority-guaranteed classes in an OBS network, most of the existing QoS schemes ignore the performance of the best-effort class. The schemes addressed this issue, such as EDS and EDS-DWG, cannot use bandwidth efficiently. As a result, the performance of the best-effort class is not good enough.

This paper first introduces an absolute QoS differentiation scheme called reserve-and-preemption (RPS), which outperforms the existing schemes such as EDS and EDS-DWG in terms of best-effort class loss rate and bandwidth utilization. To support multiple priority-guaranteed classes, an extended reserve-and-preemption scheme (RPS) is presented as well. Simulation results show that the ERPS can guarantee the loss rates of the priority-guaranteed classes while provision the best-effort class as much bandwidth as possible, and thus lower down the loss rate of the best-effort class.  When the traffic load is heavy, the ERPS can guarantee that at least the loss rate of the highest priority class can be kept under its threshold.

Alain C. Houle, GEGI, Université de Sherbrooke
Brigitte Jaumard, CIISE, Concordia University

Ammar Metnani, DIRO, Université de Montréal 

We present different scenarios based on a dynamic grooming policy for WDM mesh networks. it goes from a first scenario where no perturbation is allowed for the users to a scenario where we allow to reset grooming, routing and wavelength assignment in order to minimize the blocking rate and the costs. Clearly, a compromise has to be found taking into account the scalability of the selected solution scheme, and its resulting cost.

Photonic Millimeter-Wave Generation Using Optical Frequency Multiplication in Radio-over-Fiber Systems

Mohmoud Mohamed, Department of Electrical and Computer Engineering, Concordia University

Xiupu Zhang, Department of Electrical and Computer Engineering, Concordia University

Photonic millimeter-wave (mm-wave) generation using optical frequency multiplication (OFM) has potential to provide a cost effective solution in radio over fiber systems. As an example, we consider 30 GHz mm-wave generation using OFM. We investigate the impact of the operating biasing of an optical dual electrode driven Mach-Zehnder modulator (DE-MZM). Our results show that 30-GHz mm-wave generation using a third-order harmonic with the minimum transmission bias of the DE-MZM will lead to a better system performance compared to the maximum transmission biasing. In other words, an optimum sinusoidal RF exists for a given mm-wave.

Compressed BER monitoring in all-optical networks

Yvan Pointurier, Department of Electrical and Computer Engineering, McGill University

Mark Coates, Department of Electrical and Computer Engineering, McGill University

We tackle the problem of bit-error rate (BER) monitoring in all-optical networks (networks with no electrical regenerators) impaired by  amplifier noise.  Because BER monitoring devices are expensive, it is desirable to limit the number of deployed devices in an all-optical network, thereby limiting the number of lightpaths directly monitored at any time.  However, it is also desirable to know the quality of service and hence the BER of all lightpaths established in the network, at all times.  In this poster, we tackle the problem of estimation of the BER of the non-observed lightpaths, given a few lightpaths are directly monitored.

To estimate the missing BERs, we apply ``compressed sensing'' techniques.  We rely on the knowledge of the underlying topology of the network, which is assumed to be known by the network administrator; since multiple paths share the same links, there exists a space correlation between the lightpaths' BERs, which we exploit to estimate the unobserved BERs.  Moreover, BERs tend to be slow-varying, and hence there is also a time-correlation between the lightpaths' BER, which we exploit as well.  Simulation results show that it is possible to have a good view on the whole network behavior, more specifically, the BER on any lightpath at any time, using only a small number of optical monitoring devices.

An Improved Approach for Enhanced Availability Analysis and Design Methods in p-Cycle-Based Networks

Amin Ranjbar, CIISE, Concordia University

Chadi Assi, CIISE, Concordia University

We study the availability of the end-to-end traffic in p-cycle based mesh networks designed to protect against single link failures. The p-cycle length as well as its topology plays a vital role in determining the availability of the spans(s) protected by the p-cycle. We derive the relationship between the unavailability of a span(s) and the topology of the p-cycle allocated for the restoration of the span(s). We try to design a method for allocating p-cycles such that the end-to-end unavailability is bounded by upper limit and upper limit can be varied as desired based on the fact that the end-to-end unavailability of a working path depends not only on the length of the restoration path but also on the number of the spans along working path. Results show that more capacity is required to guarantee a lower end-to-end availability. Results also show that shorter service paths tend to use longer p-cycles that longer service paths to obtain same level of the availability; this is expected since the path length, apart from the p-cycle length, also plays a role in determining the availability of the service path. We find optimized formulation by using joint optimization of working path routing and p-cycle placement. In joint optimization, we consider the routing of the demands with the placement of the p-cycles for the protection of those demands simultaneously to find optimized formulation. We compare this formulation with the formulation, which rather limits the hop count of the candidate p-cycles to provide a lower end-to-end unavailability. We notice that directly limiting the end-to-end unavailability and using joint optimization of working path routing and p-cycle placement give better results in terms of spare capacity redundancy than limiting the hop count of p-cycles. This is because shorter paths my sometimes have less availability than longer one, therefore a better utilization of the achieved working path and allocated spare capacity.

Equipment Allocation in Video-on-Demand Network Deployments

Frederic Thouin, Department of Electrical and Computer Engineering, McGill University

Mark Coates, Department of Electrical and Computer Engineering, McGill University

Video-on-Demand (VoD) services are very user friendly, but also complex and resource demanding. Deployments involve careful design of many mechanisms where content attributes and usage models should be taken into account. We propose a methodology to solve, and define, the VoD Equipment Allocation Problem of determining the number and type of streaming servers with directly attached storage (VoD servers) to install at each potential location in a metropolitan area network topology such that deployment costs are minimized. We develop a cost model for VoD deployments based on streaming, storage and transport costs and train a parametric function that maps the amount of available storage to a worst-case hit ratio. We observe the impact of having to determine the amount of storage and streaming co-jointly, and determine the minimum demand required to deploy replicas as well as the average hit ratio at each location. Because the amount of storage required, to achieve a given hit ratio, is often much smaller than the streaming requirements unnecessary storage is installed, leading to higher deployment cost.

A Novel Approach for Overlapped Segment Shared Protection in Multi-domain Networks
Brigitte Jaumard, CIISE, Concordia University

Linh Truong, DIRO, Université de Montréal 

Routing for Overlapped Segment Shared Protection in multi-domain networks is more difficult than that of single domain networks because of the scalability requirement. In previous works, we have introduced some solutions for this problem. In this paper, a novel routing approach is proposed. The main idea is that some intra-domain paths are pre-selected for carrying working and backup traffic between domain border nodes in favoring future backup bandwidth sharing. These intra-paths will be handle as single links. Routing will be performed in intra-path level without going down unnecessarily to physical link. The novel approach benefits of good scalability and exact routing thanks to the pre-selection of intra-paths. The experimental results show also that the quality of the new approach is close to that of optimal solution in single-domain networks and outperforms those of previously proposed solutions.

Brigitte Jaumard, CIISE, Concordia University

François Vanderbeck, Université de Bordeaux

Benoit Vignac, DIRO, Université de Montréal